Google introduced an API that picks an answer from predefined options in a blog post on 11 November. The Gemini Decisions API runs on the new Gemini Decision 1 model. Google lists uses such as choosing a model or an AI agent’s tool, and rule checks.

A conventional language model generates its answer token by token. Gemini Decision 1 scores the options a developer supplies in a single forward pass and returns a probability for each. Every response includes a confidence score between 0 and 1, so a program can route uncertain cases to a fallback system or a human.

A request can hold one input of up to 256,000 tokens and up to 1,000 questions. The input can contain text, images, audio, video and PDFs. A question can be a yes/no question (noul), a multiple-choice question with 2–255 options or a rating on an ordinal scale.

Google said median latency is 150–250 milliseconds for inputs of up to 16,000 tokens. The company said the time stays almost the same with 1 to 100 questions.

Lowest calibration error in Google’s comparison

Google compared the model with two previously released decision models, TypeSafe AI’s Jev 1.14 and OpenAI’s GPT-6 Luna, and with its own Gemini 3.5 Flash-Lite. In Google’s table, as cited by MarkTechPost, Gemini Decision 1 leads seven of nine accuracy benchmarks. In two of these (audio and video), Flash-Lite was the only comparison model. Gemini Decision 1 has the lowest calibration error. This metric shows how far a model’s stated confidence differs from its actual accuracy.

Source: Google’s table, as reported by MarkTechPost. Refusals and safety blocks are counted as wrong answers.

The figures come from Google and cover benchmarks the company chose. Jev was the cheapest on the shared benchmarks. Google’s claim that the model is nearly five times cheaper refers to the comparison with Flash-Lite.

Blogger Simon Willison tested the model on a personal video. Jev and OpenAI return 1.0 or 0.0 for a yes/no question, Willison wrote, whereas Gemini gave, for example, about 0.92 for a question about whether a robot appears. MarkTechPost reported that the documentation limits the yes probability to between 0.01 and 0.99.

Price and availability

Google does not bill output tokens. The shared input is billed once per request, and each question adds only its own instruction and criteria tokens. MarkTechPost reported that there is no caching or Batch API yet, so repeated input is paid for every time. Jev costs $0.042 per 1 million input tokens.

The model cannot be downloaded to run locally, MarkTechPost said. Developers use it through Google’s Gemini API (v1beta, google-genai SDK 3.2 or later) and can try it free in Google AI Studio. Support is coming to OpenRouter, LangChain and other services, and Google said it would reach Google Cloud soon.

What this is based on

1. Gemini Decisions API: Fast, calibrated multimodal decisions

   Publisher: blog.google · published 11 November 2026 · retrieved 12 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

2. Gemini Decisions API

   Publisher: simonwillison.net · published 12 November 2026 at 00:03 (UTC+02:00) · retrieved 12 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

3. Gemini 4 Carbon is now available

   Publisher: blog.google · published 10 November 2026 · retrieved 11 November 2026 at 05:11 (UTC+02:00) · discovery channel: web

4. Introducing Gemini for healthcare | Google Cloud Blog

   Publisher: cloud.google.com · published 10 November 2026 · retrieved 12 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

5. Google Launches Gemini Decisions API With Gemini Decision 1 at $0.10/M Tokens

   Publisher: www.marktechpost.com · published 11 November 2026 at 23:21 (UTC+02:00) · retrieved 12 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

6. Google joins the decision-model wave with Gemini Decision 1, scoring 1,000 questions in a single pass

   Publisher: the-decoder.com · published 11 November 2026 at 22:12 (UTC+02:00) · retrieved 12 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

7. Google、選択肢から答えを選ぶことに特化した「Gemini Decisions API」発表 回答を生成せず確率で返す

   Publisher: www.itmedia.co.jp · published 12 November 2026 at 09:03 (UTC+02:00) · retrieved 12 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

AWS released AWS Emulator on Wednesday 11 November. The same day, Matthew Goldberg, the AWS engineer who led the project, described on an Amazon blog how agents built it. Autonomous agents wrote the whole emulator, Goldberg said: ‘Our team hasn’t written or reviewed any of the code.’

The emulator is a local program that responds to AWS API calls. When the AWS CLI or SDK is pointed at it, an application’s requests go to the emulator rather than to AWS. It can run on a developer’s own machine, on a development server or in a CI pipeline for automated testing.

GeekWire reported that the emulator covers more than 200 AWS services. Goldberg said it is built on 438 AWS API models. ITmedia, however, reported that the depth of implementation varies by service and operation. With each release, AWS publishes per-service fidelity scores, which show how closely the emulator mimics AWS.

Three agent teams and a review gate

Goldberg said three teams of agents build the emulator. The first writes tests from public API models and checks that each one works on real AWS. A test might, for example, create a queue, send a message and delete the queue. The second team runs the tests against the emulator and logs failures to be fixed. In the third team, each agent fixes one issue.

A change reaches the main branch only once the code compiles. Reviewing agents must also confirm that it follows coding conventions, is secure and does not reduce fidelity to AWS. Hundreds of agents fix thousands of issues a day, Goldberg said. Simon Willison wrote in a blog post that the project accumulated 29,684 commits in its first three weeks.

The project began as a weekend experiment by Goldberg. It was inspired by an experiment by Anthropic researchers in which agents built a C compiler autonomously. Goldberg said Amazon’s Kiro coding agent produced a viable first version of the Amazon SQS queue service in 15 minutes.

Not for production or security testing

AWS advises against using the emulator in production or for security testing. ITmedia reported that the emulator does not handle IAM and resource policies, which define access permissions, the same way AWS does. Resources are held in memory and disappear when the emulator is stopped.

GeekWire reported that IDC analyst Devin Pratt told InfoWorld the emulator does not yet mimic request throttling, outages or regional differences in the availability of services and features. Pratt also said it does not isolate the code it runs. Flaws the agents have not yet fixed are publicly listed in the repository’s issues/things-that-do-not-work directory.

Availability and setup

The emulator needs no AWS account, carries no extra charge, and its code is on GitHub, mostly under a CC0 licence. GeekWire reported that LocalStack, the best-known alternative, this year limited its free version to non-commercial use and began requiring its users to have an account.

On Linux, the emulator can be installed from an RPM or Debian package; on macOS and Windows it runs in a Docker container. Telemetry, which sends usage data to AWS, is on by default. Instructions for turning it off and for other setup are in the README file on GitHub and in the AWS user guide.

What this is based on

1. aws-emulator

   Publisher: simonwillison.net · published 11 November 2026 at 23:18 (UTC+02:00) · retrieved 12 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

2. How we built the agents that built AWS Emulator

   Publisher: www.aboutamazon.com · published 11 November 2026 at 19:25 (UTC+02:00) · retrieved 12 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

3. github.com

   Publisher: github.com · source publication time unknown · retrieved 12 November 2026 at 05:24 (UTC+02:00) · discovery channel: GitHub

4. AWS gives developers a free local stand-in for its cloud, built from scratch by AI agents

   Publisher: www.geekwire.com · published 11 November 2026 at 20:51 (UTC+02:00) · retrieved 12 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

5. Amazon、AWSのAPIを手元で再現する「AWS Emulator」を無償公開 コードはすべてAIエージェントが記述

   Publisher: www.itmedia.co.jp · published 12 November 2026 at 11:56 (UTC+02:00) · retrieved 12 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

Reuters reported on 11 November, citing Chinese procurement records, that Xi’an-based Northwestern Polytechnical University paid 874,700 yuan (about $123,000) for two Nvidia H200 chips in October. The seller was the Chinese company Shaanxi Bosuan Network Technology, not Nvidia.

The H200 is among Nvidia’s most powerful chips for training and running AI models. The university researches military aircraft, missiles and satellites. In 2001 the US Commerce Department added it to the Entity List, whose members are in practice barred from controlled US technology. In 2019 the department tightened the restrictions so that licence applications are generally rejected. In January the Trump administration began allowing H200 exports to China case by case, but the decision did not cover entities on the list.

Chips bought and rented through intermediaries

This year the university also bought computing capacity from Chinese companies under five contracts. In January its cybersecurity school signed two seven-month contracts for access to servers containing eight H200 chips each. The university also rented access to servers with A100 and A800 chips, which may not be exported to any Chinese entity. The January licence conditions require safeguards that stop listed and military users from accessing chips exported under licence remotely. The records do not, however, show where the chips came from or where the rented servers are located.

Reuters reviewed 26 records from this year. According to these records, other universities, including Shenzhen University and defence-linked Beihang, have also bought restricted chips from domestic suppliers. Northwestern Polytechnical, Nvidia and the Commerce Department’s export control agency, BIS, did not respond to Reuters’ questions.

Commerce secretary holds back, Congress weighs legislation

The report came amid a scandal: Nvidia chief executive Jensen Huang has admitted a relationship with Alice Huang, an executive at Nvidia customer Megaspeed. CNBC reported that the New York Times wrote in October that US officials regard Alice Huang as a source for Chinese intelligence. China has denied the spying allegations.

Axios reported on 11 November, citing six sources, that President Donald Trump has approved action at least three times this year against companies suspected of diverting Nvidia chips to China. Commerce Secretary Howard Lutnick has delayed implementation and warned that the restrictions ‘will start a trade war with China’, the sources said. A Commerce Department spokesperson denied that Treasury Secretary Scott Bessent had rebuked Lutnick in a meeting and did not immediately comment on the other claims.

Germany’s heise, citing The Information, reported that Alice Huang’s family network of companies supplied at least $7.5 billion worth of Nvidia chips to China. The Information’s sources are two people who know Alice Huang.

Members of Congress, aides and lobbyists interviewed by Politico expect the scandal to make Congress more likely to tighten chip exports in this year’s defence budget bill. Three lobbyists said the most likely addition is the Chip Security Act, which would require advanced chips to carry a mechanism that verifies their location. Reuters reported that three Republican committee chairs have asked Jensen Huang to explain these connections by 18 November.

What this is based on

1. Scoop: Lutnick stalled Trump's Nvidia–China chip curbs after Oval talks

   Publisher: www.axios.com · published 11 November 2026 at 22:59 (UTC+02:00) · retrieved 12 November 2026 at 05:25 (UTC+02:00) · discovery channel: web

2. Nvidia loses ground in Washington chip fight after CEO spy scandal

   Publisher: www.politico.com · published 11 November 2026 at 05:18 (UTC+02:00) · retrieved 12 November 2026 at 05:26 (UTC+02:00) · discovery channel: web

3. Bericht: Jensen Huangs Affäre brachte Nvidia-Chips im Milliardenwert nach China

   Publisher: www.heise.de · published 11 November 2026 at 06:03 (UTC+02:00) · retrieved 12 November 2026 at 05:26 (UTC+02:00) · discovery channel: web

4. Nvidia sells off for second day in fallout from Jensen Huang’s affair with alleged spy

   Publisher: www.cnbc.com · published 10 November 2026 at 11:14 (UTC+02:00) · retrieved 12 November 2026 at 05:21 (UTC+02:00) · discovery channel: web

5. Wow: the @WSJ says Nvidia's board is preparing to oust Jensen Huang https://t.co/jgYCHCzWrb

   CaseyNewton · Publisher: x.com · published 12 November 2026 at 00:35 (UTC+02:00) · retrieved 12 November 2026 at 04:10 (UTC+02:00) · discovery channel: X

6. On today's @MachineGodsPod, Kevin and I discuss whether Jensen Huang survives as Nvidia CEO. Plus: a great interview with Rep. Greg Casar on the incoming Democratic majority and how it plans to…

   CaseyNewton · Publisher: x.com · published 11 November 2026 at 19:17 (UTC+02:00) · retrieved 12 November 2026 at 04:10 (UTC+02:00) · discovery channel: X

7. Exclusive-PLA-linked Chinese university bought Nvidia H200 chips despite US blacklist, records show

   Publisher: www.internazionale.it · published 11 November 2026 at 07:30 (UTC+02:00) · retrieved 12 November 2026 at 05:27 (UTC+02:00) · discovery channel: web

Evaluation company Artificial Analysis published a comparison on X on 12 November in which it ran Anthropic’s current Claude models at every reasoning effort level. It set each setting’s score on its own Intelligence Index benchmark suite against the cost of one task. The company concluded that moving up a model tier buys more intelligence per dollar than raising a smaller model’s reasoning effort to the maximum.

Reasoning effort is an API setting, called effort. Anthropic’s guidance says it governs how much the model thinks, calls tools and checks its own work. On the Messages API the default is high on most models but medium on Opus 5.5 and Haiku 5.5. From cheapest to most expensive, Anthropic’s models are Haiku 5.5, Sonnet 5.5, Opus 5.5 and the frontier model Fable 5.2.

A larger model beat a smaller one at maximum effort

Artificial Analysis said the largest model sits on the efficiency frontier at almost every cost level, meaning no other setting scores more for the same price.

Opus 5.5 scored 57 at xhigh and Sonnet 5.5 scored 55 at max. Opus’s cost per task was 40 per cent lower, and it took about half as long as Sonnet. The chart in the post shows 15.9 and 32.6 minutes per task.

Fable 5.2 at high beat Opus 5.5’s max score, but the savings were smaller. The chart shows Fable scoring 3 points more at 10 per cent lower cost and taking 16.7 minutes per task against Opus’s 25.1. Fable 5.2 at low also beat Opus 5.5 at high.

The price list would not predict the result: Anthropic says each of Opus 5.5’s token prices is twice Sonnet 5.5’s. Anthropic’s guidance notes in general that a more capable model can complete a task with less work.

The rule does not hold in the cheapest tier

Artificial Analysis said every step up from Haiku 5.5 at max costs more. In the chart, Sonnet 5.5’s cheapest setting, low, scores lower than Haiku 5.5 at max yet still costs more.

The figures apply to Artificial Analysis’s own task set. Anthropic’s guidance says how much accuracy a lower reasoning effort costs depends on the task. In Anthropic’s own test on a subset of the SWE-bench Pro coding benchmark, Opus 5.5 at xhigh matched Fable 5.2 at high for a little over half the cost per solved task. For most agent tasks, the company recommends starting with Opus 5.5 at its default effort and comparing models on your own workload by what a finished task costs. If quality falls short, the guidance advises restoring the lowered reasoning effort or trying the next model tier at its default effort.

What this is based on

1. With the current Claude lineup, moving up a model tier beats maxing out effort to achieve the highest intelligence per dollar At nearly every Cost per Task, we see the largest model sits on the…

   ArtificialAnlys · Publisher: x.com · published 12 November 2026 at 02:18 (UTC+02:00) · retrieved 12 November 2026 at 04:15 (UTC+02:00) · discovery channel: X

2. Optimizing for cost and intelligence

   Publisher: platform.claude.com · source publication time unknown · retrieved 12 November 2026 at 05:31 (UTC+02:00) · discovery channel: web

Benchmarking firm Artificial Analysis published a comparison on X on 12 November in which it ran the same scientific-image benchmark in two ways. Models either used tools as part of an agent or answered a single request without tools. Tools raised models’ scores from about 25 per cent to about 47 per cent on average, the company said.

What the benchmark measures

VIALS-AA is Artificial Analysis’s implementation of Handshake AI’s VIALS benchmark. In its 161 tasks, a model interprets images used in life-sciences work, such as blot images, microscopy images, flow cytometry plots, plasmid maps and phylogenetic trees. The VIALS research paper has been published as a preprint on arXiv.

By default, models run in the Stirrup harness. A harness is a program that controls an agent: it gives the model tools and runs its steps. The model can run Python code in a sandbox with no network access and open images with a separate tool. In the comparison run, the model received the question and images in a single request without tools.

Each task is run three times, and the language model GPT-6 Luna grades the answers. The score (pass^3) is the share of tasks the model answered correctly in all three runs. A refusal or a missing answer scores zero.

Results

Claude Fable 5.2 (max) scored 67 per cent with tools and 44 per cent without. Claude Haiku 5.5 (max) reached 49 per cent with tools, higher than Fable without tools. Of the models in the published chart, the best without tools was GPT-6 Astra (max), at 50 per cent.

The figures show the effect of tools in this image-interpretation benchmark, not which models are better in general.

Costs

Tools raised every model’s cost per task, but by different amounts. Fable 5.2’s cost rose from $0.73 to $4.20, a 5.8-fold increase; GPT-6 Astra’s from $0.21 to $0.27; and GPT-6.1 Sol’s from $0.05 to $0.06. Costs grew most for Qwen3.8 Max (28-fold), Gemini 3.8 Flash (20-fold) and MiMo-V2.6-Pro (15-fold). Artificial Analysis calculated costs from token counts and model prices, assuming a typical share of cache hits.

Artificial Analysis concludes that, when interpreting scientific images, a cheaper model in a harness can beat a large model answering a single request on both price and score. In the company’s example, Haiku with tools scores 49 per cent at $0.07 per task, while Claude Opus 5.5 without tools scores 34 per cent at $0.33.

What this is based on

1. VIALS-AA Benchmark | Artificial Analysis

   Publisher: artificialanalysis.ai · source publication time unknown · retrieved 12 November 2026 at 05:34 (UTC+02:00) · discovery channel: web

2. Take away its tools, and Claude Fable 5.2 scores below Claude Haiku 5.5 with tools on VIALS-AA; across models tool use nearly doubles scores, from ~25% to ~47% on average VIALS-AA is our…

   ArtificialAnlys · Publisher: x.com · published 12 November 2026 at 01:46 (UTC+02:00) · retrieved 12 November 2026 at 04:15 (UTC+02:00) · discovery channel: X

3. Tools raise the cost per task for every model, but by very different amounts. Astra costs $0.27 per task with tools and $0.21 without, and Sol $0.06 and $0.05. Fable costs $4.20 with tools and $0.73…

   ArtificialAnlys · Publisher: x.com · published 12 November 2026 at 01:46 (UTC+02:00) · retrieved 12 November 2026 at 04:15 (UTC+02:00) · discovery channel: X

4. Explore the results of VIALS-AA (with tools) and compare models on Score, Cost per Task and Tool Use: https://t.co/4zhMuS3v2Q

   ArtificialAnlys · Publisher: x.com · published 12 November 2026 at 01:46 (UTC+02:00) · retrieved 12 November 2026 at 04:15 (UTC+02:00) · discovery channel: X

5. Claude Fable 5.2 is the new leader on AA-Robotics v0.1, scoring 75, ahead of GPT-6 Astra (69) and GPT-6.1 Sol (65) Fable 5.2 is the first model to achieve a score on the measure box task. It did so…

   ArtificialAnlys · Publisher: x.com · published 12 November 2026 at 04:00 (UTC+02:00) · retrieved 12 November 2026 at 04:15 (UTC+02:00) · discovery channel: X

6. 会社員の仕事に「Claude Fable 5.2」は必要か？ 調査、要約、資料作成でOpus／Sonnetと比較してみた

   Publisher: www.itmedia.co.jp · published 12 November 2026 at 10:00 (UTC+02:00) · retrieved 12 November 2026 at 05:34 (UTC+02:00) · discovery channel: web

7. Google has released its first decision model, Gemini Decision 1, scoring 66 on the Artificial Analysis Decision Index and leading in knowledge and reasoning benchmarks Gemini Decision 1 from…

   ArtificialAnlys · Publisher: x.com · published 11 November 2026 at 20:18 (UTC+02:00) · retrieved 12 November 2026 at 04:15 (UTC+02:00) · discovery channel: X

On 10 November, AWS published a security bulletin and a GitHub security advisory about a flaw in the AWS API MCP Server’s custom security policies. The flaw affects versions 0.2.13 to 1.5.0.

AWS API MCP Server is an open-source MCP server that users run themselves. MCP is a protocol through which an AI assistant uses external tools. Here, the assistant manages an AWS account through AWS API calls. The operator can define a custom security policy for the server. Operations listed in its denyList are blocked, and those in its elicitList require user confirmation before they run.

Why the rules failed

The server normalises operation names into a consistent format before comparing them with the policy. AWS said the server normalised names containing consecutive capital letters incorrectly. As a result, even a correctly written denyList or elicitList entry never matched such an operation.

The CVE description says the flaw could let through operations across multiple AWS services. Examples in the GitHub advisory include creating a federated identity provider, deactivating an MFA (multi-factor authentication) device and deleting database clusters.

Who is affected

The flaw affects only a locally run server whose custom policy lists affected operations. The user must also be using an AI assistant. The AWS API MCP Server that AWS offers as a remote service is not vulnerable. Nor does the flaw affect the READ_OPERATIONS_ONLY and REQUIRE_MUTATION_CONSENT settings.

The CVE record and the GitHub advisory rate the flaw High, with a CVSS 4.0 score of 8.5. AWS’s own bulletin classes it as Important.

Upgrading is the only fix

AWS said only upgrading to version 1.5.1 or later fixes the flaw, and there is no workaround. The company urges users to patch forks and other derivatives of the server as well.

Japanese technology site Tech InnovateIT reminds readers that the MCP server must be restarted after the upgrade. Restricting IAM permissions adds protection but does not replace the upgrade, the site says. It advises testing deny rules and confirmation prompts in a development or test environment, not in a production account.

Tech InnovateIT reported that the latest version on PyPI on 11 November was 1.6.0, which also fixes a separate missing OAuth check. The developer’s separate security advisory, the site wrote, advises setting the expected token audience or client ID when version 1.6.0 is used with streamable-http transport and OAuth. The setting is not needed in every configuration, for example with stdio.

The site also reported that PyPI directs users to the AWS MCP Server, named as the successor. Citing the developer’s migration guide, it said development of new features ended on 15 July 2026 and the server is due to be removed on 15 July 2027.

What this is based on

1. CVE-2026-125271 - AWS API MCP Server Security Policy Bypass via Incorrect Operation Name Normalization

   Publisher: aws.amazon.com · source publication time unknown · retrieved 12 November 2026 at 05:29 (UTC+02:00) · discovery channel: web

2. CVE Record: CVE-2026-125271

   Publisher: www.cve.org · source publication time unknown · retrieved 12 November 2026 at 05:29 (UTC+02:00) · discovery channel: web

3. github.com

   Publisher: github.com · published 10 November 2026 · retrieved 12 November 2026 at 05:29 (UTC+02:00) · discovery channel: GitHub

4. NVD-CVE-2026-125271

   Publisher: nvd.nist.gov · source publication time unknown · retrieved 12 November 2026 at 05:30 (UTC+02:00) · discovery channel: web

5. CVE-2026-125271: AWS API MCP Server Security Policy Bypass via Incorrect Operation Name Norm

   Publisher: casco.com · published 10 November 2026 · retrieved 12 November 2026 at 05:29 (UTC+02:00) · discovery channel: web

6. denyList不備を修正｜AWS API MCP Serverを1.5.1へ

   Publisher: techinnovateit.com · published 11 November 2026 at 11:49 (UTC+02:00) · retrieved 12 November 2026 at 05:29 (UTC+02:00) · discovery channel: web

Daniel Lenton of unify published a preprint on arXiv on 10 November describing the Continual-ARC benchmark. The benchmark measures whether an AI agent saves effort on recurring tasks. The study has not been peer reviewed.

The benchmark uses visual puzzles from ARC-AGI-1. The agent has to infer, from a few examples, a rule that turns one grid into another. A task sequence contains 200 task instances, and 25 rules return after gaps, each time with new input. An extra example and a wrong answer each cost one unit; an unsolved task costs the full budget of 16 units.

At the end, the frozen agent solves new versions of familiar tasks and 100 unseen tasks. The former measure memory, the latter generalisation. The score is the fraction saved compared with the same agent whose memory is wiped before every task.

Harnesses remembered only when instructed

A harness is a program that controls an agent: it gives the model tools and runs its steps. Lenton ran three harnesses on GPT-6 Luna with their own memory features switched on: Prime Agent, Hermes Agent and OpenClaw. Memory produced no measurable savings at any reasoning effort.

Prime Agent kept an entry for later sessions only if the agent marked it as global. The agent never did. Hermes Agent’s skills contained general advice that did not reduce spending. OpenClaw carried nothing over to persistent memory.

When prompted to save liberally, only Prime Agent saved measurably: 0.42 on familiar tasks. The harnesses saved 0.62–0.72 on familiar tasks when their instructions told them to write a Python program for each task, check it against the examples and reuse it. Lenton’s reading is that the harnesses were capable of a good result, but neither the model’s judgement nor the harnesses’ default prompts steered the agent to save the right things.

Training the weights helped on new tasks

No harness-based method measurably improved results on unseen tasks. During the task sequence, Lenton trained the weights of the open-weight model Qwen3.5-9B to imitate GLM-5.3 Flash. With this method, DAgger, the model saved 0.29 on new tasks.

The table shows Qwen’s results with each method.

Fraction saved compared with the empty-memory version; higher is better. The program library stores a program checked against the examples for each task. Source: Continual-ARC preprint, ARC-AGI-1.

Apart from training, the figures for new tasks do not differ statistically from zero, and nor do the Prime Agent rows, even on familiar tasks.

Part of the gain from training came from learning to use the harness. Untrained, Qwen solved 18 of 100 new tasks under Prime Agent and 31 in a simple loop without tools. The trained model solved 47–52, so it also learned from the tasks themselves, Lenton said. DAgger needs a stronger teacher model whose reasoning tokens are available.

Lenton wrote on X that the harnesses remembered nothing of value. The preprint limits this finding to harnesses run with default settings. Most runs used relatively small models and the easiest of the three ARC tracks. How harnesses learn in other domains remains open, Lenton said. Lenton’s employer, unify, is itself developing continual learning based on weight training.

The code is on GitHub under the Apache-2.0 licence, and the run logs are under CC BY 4.0. Anyone can have their own solution checked against unify’s secret task set by paying for the model calls in advance.

What this is based on

1. Continual-ARC: Measuring Skill Acquisition over a Working Lifetime

   Lenton, Daniel · Publisher: arxiv.org · published 10 November 2026 at 20:55 (UTC+02:00) · retrieved 12 November 2026 at 05:35 (UTC+02:00) · discovery channel: web

2. github.com

   Publisher: github.com · source publication time unknown · retrieved 12 November 2026 at 05:35 (UTC+02:00) · discovery channel: GitHub

3. x.com

   Daniel Lenton · Publisher: x.com · published 11 November 2026 · retrieved 12 November 2026 at 05:34 (UTC+02:00) · discovery channel: X

United Airlines announced the decision in a press release on Wednesday 11 November. The ban covers third-party personal AI agents that the company has not authorised. An AI agent is a program that uses a language model to handle tasks on a user’s behalf. CBS News and Scripps News give Meta’s Muse and OpenAI’s Dots as examples.

United is restricting agents’ access through its own online services as well as through travel agencies and aggregators. It is also updating its booking, ticketing and distribution terms. ITmedia reported that United did not say which agents the restriction covers or when it will take effect.

Agents can still show users United flights but cannot complete the purchase, The Verge reported. Henry Harteveldt, an aviation analyst at Atmosphere Research Group, told CBS News that travellers can still book United flights themselves through aggregators such as Kayak, Hopper and Google Flights.

United cites security and aviation rules

The press release said sensitive information could be exposed if customers give their login details to an agent. Agents’ actions could also lead to unintended purchases, account changes or use of miles from the MileagePlus loyalty programme.

United also claims that automated actions could breach rules set by the US Federal Aviation Administration (FAA) and the Department of Transportation (DOT). In the company’s example, a person must confirm at check-in that bags contain no banned substances such as explosives.

The Verge, citing The Wall Street Journal, reported that United has not yet seen many people try to book flights using agents.

Analyst: customer relationships are also at stake

The Verge and Engadget quote The Wall Street Journal, which reported that chief executive Scott Kirby said at an event on Tuesday: ‘To be honest, I don’t want to create another middleman.’

Harteveldt said United does not want a new, possibly even more powerful middleman alongside search engines, comparison sites and online travel agencies. Flight searches cost airlines money, Harteveldt noted. In Harteveldt’s view, United clearly wants to avoid the cost of a rush of searches by AI bots.

Not all airlines are taking the same approach. CBS reported that Southwest has made its flights visible to ChatGPT users. The aviation site Paddle Your Own Kanoo wrote that Delta reportedly already blocks third-party agents.

ITmedia reported that companies in other sectors have also restricted third-party agents since Meta released Muse in September. Amazon has blocked purchases made by Muse since 20 September. Between 13 and 20 October, JPMorgan Chase, Charles Schwab, Citigroup and Wells Fargo acted against agents that log in with customers’ own credentials. The banks either blocked the agents or shifted the risk to customers through their agreements.

United says it is updating its Digital Trust Center security site. After the update, the site will provide information on the restrictions and on protecting accounts, as well as guidance on getting help if an agent takes an unintended action. The company added that it will continue to assess how agents operate on its services as they develop.

What this is based on

1. United Airlines bans AI agents from booking flights on behalf of travelers

   Publisher: www.cbsnews.com · published 11 November 2026 at 21:37 (UTC+02:00) · retrieved 12 November 2026 at 05:35 (UTC+02:00) · discovery channel: web

2. United Airlines is Blocking AI Booking Agents: They Might Breach FAA And DOT Regulations

   Publisher: www.paddleyourownkanoo.com · published 11 November 2026 at 22:08 (UTC+02:00) · retrieved 12 November 2026 at 05:36 (UTC+02:00) · discovery channel: web

3. United won’t let your AI agent book your flight

   Publisher: www.theverge.com · published 11 November 2026 at 23:54 (UTC+02:00) · retrieved 12 November 2026 at 05:36 (UTC+02:00) · discovery channel: web

4. ユナイテッド航空、外部のAIアシスタントによる航空券の予約などを禁止へ 「チェックインには人間の確認が必要」

   Publisher: www.itmedia.co.jp · published 12 November 2026 at 12:49 (UTC+02:00) · retrieved 12 November 2026 at 05:36 (UTC+02:00) · discovery channel: web

5. United Restricts Access for Third-Party AI Agents

   Publisher: united.mediaroom.com · published 11 November 2026 · retrieved 12 November 2026 at 05:37 (UTC+02:00) · discovery channel: web

6. United Airlines blocks AI assistants from booking flights

   Publisher: www.scrippsnews.com · published 11 November 2026 · retrieved 12 November 2026 at 05:38 (UTC+02:00) · discovery channel: web

7. United Is Restricting AI Agents From Booking Flights

   Publisher: tech.yahoo.com · published 11 November 2026 at 22:09 (UTC+02:00) · retrieved 12 November 2026 at 05:38 (UTC+02:00) · discovery channel: web

8. Exclusive | United Airlines Is Banning AI Bots From Booking Flights

   Publisher: www.wsj.com · source publication time unknown · retrieved 12 November 2026 at 05:35 (UTC+02:00) · discovery channel: web

Bloomberg reported the lawsuit in an article that Malaysia’s The Star published on 11 November. The quantitative trading unit of Steve Cohen’s investment firm Point72 brought the suit. Quantitative trading is trading based on mathematical models.

Cubist Systematic Strategies sued its former research analyst Xin Jin on 9 November in state court in Stamford, Connecticut. This account draws on court documents cited by Bloomberg and sets out Cubist’s allegations, on which the court has not ruled.

Jin joined the firm in 2019. The suit says Jin was the lead research analyst in KEPL, a 14-person team specialising in statistical arbitrage, which uses statistical models to find price anomalies.

What Cubist alleges

The suit alleges that Jin used Anthropic’s Claude and OpenAI’s Codex, both available at work, to systematically review, analyse and distil the KEPL team’s research results and source code into a readily extractable format. Cubist said the material could have been used to rebuild the team’s work outside the firm’s systems. The firm also alleges that Jin moved data to personal devices and deleted the AI tool history to evade detection.

According to Cubist’s own investigation, Jin had tried to steal, and likely did steal, confidential information. Cubist describes the code as the lifeblood of its business and a secret recipe that lets it generate positive returns.

Cubist began its investigation in July after team members raised concerns about Jin’s behaviour. The suit says Jin hinted at leaving the firm. Jin also told colleagues about buying market data for personal use, spending thousands of dollars on AI-powered coding tools and considering buying a computer with a terabyte of memory.

In July Jin told human resources about considering a move to a competitor and was put on leave. Cubist fired Jin in September after Jin reported accepting a position with a competitor. The suit does not name the competitor.

Cubist is seeking damages of an undisclosed amount. It also wants the court to bar Jin from disclosing its confidential information and from competing with it. The firm wants to examine Jin’s personal devices as well, but says Jin has refused.

Jin denies the allegations

Jin’s lawyer, Rollo Baker, said Cubist’s claims are without merit. Baker said Jin looks forward to setting the record straight.

In a filing on 10 November, Jin stated having cooperated with the former employer in a proposed mediation. The disputes also include Jin’s claim to more than $30 million in investments and compensation that Jin considers owed. The filing says Jin is due to start with a new employer in June 2028. Baker did not say which company it is.

What this is based on

1. Point72 claims fired quant analyst used AI tools to steal code

   Publisher: www.thestar.com.my · published 11 November 2026 · retrieved 12 November 2026 at 05:40 (UTC+02:00) · discovery channel: web

2. Point72 Accuses Fired Quant Analyst of Using AI Tools to Steal Source Code

   Publisher: www.bloomberg.com · published 10 November 2026 at 20:38 (UTC+02:00) · retrieved 12 November 2026 at 05:39 (UTC+02:00) · discovery channel: web

On 11 November, David Gewirtz wrote in ZDNET about isolating an AI agent that runs around the clock at his home and then asking it to escape. OpenClaw is a program that runs on a language model and can execute commands and send messages.

Prisoner and warden on separate machines

Gewirtz installed OpenClaw on an old M1-based Mac Studio. The machine has internet access but cannot reach the servers in his home or his personal accounts. The agent runs under its own user account and cannot access the machine’s separate administrator account. Its language model is Z.ai’s GLM-5.2, which Gewirtz runs on Ollama Cloud for a flat fee of $20 a month.

Claude Code acts as the warden on Gewirtz’s own Mac. It connects to the agent’s machine through an encrypted SSH tunnel and runs only when Gewirtz is using it. Claude Code was originally installed on the agent’s machine, but Claude pointed out that the agent could reach Claude’s own instructions and notes there.

The agent can read its configuration files but cannot change them. They contain two key instructions: ask David before any action that reaches beyond the machine, and never claim that a failed tool call succeeded. Claude spotted in OpenClaw’s log that the agent had stopped reading the file at the default size limit and had read neither instruction. The default limit was changed to fix this. Instructions work only if the AI agrees to follow them, Gewirtz said.

Eleven attempts, no breakthrough

With Gewirtz’s permission, Claude told the agent to break into the administrator side. The agent tried to write directly to the administrator account’s files, to gain administrator privileges through several back doors and to get round the permissions with scripts, file copies and links. It also tried to open the warden’s files by pointing its browser and its web-fetching tool at local file addresses, hoping those tools had wider access than the agent itself.

All 11 attempts failed and the macOS protections held, Gewirtz said. The result applies to this set-up and to the methods the agent itself tried. Gewirtz was ready to pull the machine’s power cable and switch off the router, and he warns readers not to try this at home. Finally, Claude deleted information about the test and the administrator side from the agent’s memory.

The agent has no email access. It leaves its messages as text files in a folder, from which a Python script on the administrator side sends them only to Gewirtz’s address. Gewirtz is ‘fairly convinced’ that it is safe to keep the agent running, citing the tested constraints and regular checks on them.

The AI Understanding website summarises ZDNET’s article but names ZDNET’s Charlie Osborne as the person who ran the experiment. Gewirtz mentions Osborne in the article only because Osborne has called OpenClaw a security nightmare.

What this is based on

1. I locked down my OpenClaw agent with Claude Code - then dared it to escape - ZDNET

   Publisher: zdnet.com · published 11 November 2026 at 15:09 (UTC+02:00) · retrieved 12 November 2026 at 05:41 (UTC+02:00) · discovery channel: web

2. ZDNET Tests OpenClaw Agent with Claude Code, Dares It to Escape

   Publisher: aiunderstanding.org · published 11 November 2026 at 15:29 (UTC+02:00) · retrieved 12 November 2026 at 05:41 (UTC+02:00) · discovery channel: web

Google describes the new /swarm command in the Antigravity documentation. Antigravity is Google’s programming environment in which AI agents do development work. A notice on the documentation page says /swarm is now available in preview. Google says the command suits tasks that split clearly into separate files, packages or questions. The command also lets users have another agent review and test the changes before they see them.

The coordinator divides the work, the verifier tests it

When the user runs the command and describes the task, the current conversation becomes the team’s coordinator. The coordinator splits the task into parts that do not overlap. Each file or package, for example, can get its own agent. Every worker agent gets a name, such as @parser or @docs, and starts its own conversation with no prior context. Its brief is posted for everyone to see. The brief sets out the agent’s responsibilities, what it must not touch and the condition for finishing the work.

The coordinator usually adds one verifier (@verifier) for every two to four workers. The verifier reads the changes and runs the tests. It reports problems to the agent that wrote the code, which fixes them. The coordinator does no technical work itself. It assigns tasks, assesses the results and reports to the user.

The agents do not send each other private messages. Briefs, questions, change summaries and test results go on a shared message board. An idle agent waits without using tokens and wakes when someone mentions it. If every agent falls silent mid-task, a separate @monitor agent uses the board to call on an agent that can get the work moving again.

By default, workers use the same model as the user’s conversation. Users can request a different model for one worker, such as the verifier, or for all of them. Agents work in the user’s workspace by default, but the coordinator can give each one its own copy of the repository. Removing an agent also deletes its copy, so the coordinator has the agent merge its changes first.

The usage limit runs out fast

Google warns that a team uses up the usage limit much faster than a single conversation. Each worker is a separate agent. If an agent reaches its usage limit, it stops and does not resume on its own when the limit resets. It resumes only when a new message arrives on the board, its model is changed or the user messages it. A removed agent cannot be restored.

Google recommends /swarm for large tasks that split clearly into parts. For a hard but narrow problem, Google points users to the /boost command. For projects lasting hours to days with several milestones, it recommends /teamwork.

Google AI Pro and Google AI Ultra subscribers can use the command in the Antigravity app and the Antigravity CLI, its command-line tool. Users can follow the team’s work in the app’s Swarm panel or on the command line, and the Antigravity documentation has more detailed instructions.

What this is based on

1. Swarm (/swarm) | Google Antigravity Docs

   Publisher: antigravity.google · source publication time unknown · retrieved 12 November 2026 at 05:57 (UTC+02:00) · discovery channel: web

Alibaba’s Qwen team said on X on 11 November that its next model generation, Qwen4, is on the way. The same day, the team opened a preview of Qwen4-Max. Alibaba’s changelog describes qwen4-max-preview as a multimodal frontier model: it generates text, understands images and video, and reasons before answering. The company is aiming it at programming, office work, front-end development and complex agentic tasks.

The model supports web search, tool calls and structured output. It is compatible with the OpenAI and Anthropic APIs.

A 1 million-token context, but no public benchmark results

Citing Alibaba’s documentation, Pandaily and Xenospectrum report a context window of 1 million tokens and a maximum output of 393,216 tokens. The output limit covers both reasoning and the final answer, so long reasoning leaves less room for the answer. Reasoning is on by default, and users can adjust how much the model reasons with the reasoning effort setting.

Xenospectrum said a single request can include up to 64 videos, each up to two hours long and 2 GB in size. The publication notes that the input limit does not guarantee the model will understand a long video or context accurately.

Alibaba has not published benchmark results, a model card, a parameter count, a per-token price or the model weights, Pandaily reported. Nor has the company given a release date for the full Qwen4 generation, and the claims about the model’s capabilities are its own. AIBase earlier reported leaked benchmark figures, but Alibaba has not published them.

Trial ends on 25 November

Through the API, the model is available only with a Token Plan subscription, not on pay-as-you-go billing. Pandaily said the documentation limits the trial to the period from 10:00 on 11 November to midnight on 25 November, Beijing time, which is 18:00 on 25 November Finnish time. The model will not change during the trial and will be retired afterwards.

The sources differ on the discount. Pandaily said Qwen4-Max calls consume Token Plan usage units, or credits, at a promotional rate as low as 20 per cent of the standard rate. AIBase put the price at as low as one-tenth of the original.

A personal Token Plan costs $6–43 a month and includes 20,000–200,000 credits, Xenospectrum reported. It is intended for interactive use, for example in development environments and coding agents. Its terms prohibit bulk processing with automation scripts, permanent backends for your own applications, and batch jobs. A team subscription costs $20–180 per seat per month.

The publication said Token Plan can be bought and used only in Alibaba Cloud’s Singapore region, and its API key and endpoint do not work with the standard pay-as-you-go API.

Users can also try the model in Alibaba’s own services: QwenWork, the Qoder coding tool, the Qwen app and QwenStudio.

What this is based on

1. 模型上架与更新 - 千问AI平台

   Publisher: platform.qianwenai.com · published 11 November 2026 · retrieved 12 November 2026 at 05:31 (UTC+02:00) · discovery channel: web

2. Alibaba Opens Qwen4-Max Preview on Token Plan, With a 1M-Token Context Until November 25

   Publisher: pandaily.com · published 12 November 2026 at 04:18 (UTC+02:00) · retrieved 12 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

3. Qwen teases Qwen4: Qwen4-Max preview launches on Alibaba Token Plan, limited-time as low as 10% of the original price

   Publisher: news.aibase.com · source publication time unknown · retrieved 12 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

4. Alibaba's Qwen4-Max Preview Offers a 1M-Token Context Window, but Access Ends Nov. 25

   Publisher: xenospectrum.com · published 11 November 2026 at 08:58 (UTC+02:00) · retrieved 12 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

5. 中国AI「Qwen4」ついに登場 プレビュー版を提供開始、AlibabaのAIサービスで

   Publisher: www.itmedia.co.jp · published 11 November 2026 at 13:30 (UTC+02:00) · retrieved 12 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

6. 阿里云发布基于 AMD 第六代 EPYC 处理器的第十代 ECS 企业级实例

   凌敏 · Publisher: www.infoq.cn · published 11 November 2026 at 08:44 (UTC+02:00) · retrieved 12 November 2026 at 04:27 (UTC+02:00) · discovery channel: RSS feed

7. 模型上架与更新

   Publisher: docs.modelstudio.console.alibabacloud.com · source publication time unknown · retrieved 12 November 2026 at 05:33 (UTC+02:00) · discovery channel: web

Alibaba released a preview of Qwen Code 1.0 on 11 November, merging into it the harness from its Qoder coding product. A harness is the program that controls an agent: it gives the model its tools and runs its work steps. Pandaily’s account of Alibaba’s description is based on a summary by the Chinese site IT Home.

Qwen Code is an open-source coding agent that runs in the terminal. The project also has desktop, browser, editor and chat integrations, for example with VS Code, Zed and JetBrains editors.

The GitHub release notes say v1.0.0-preview brings together the agent capabilities of Qoder and Qwen Code: understanding code, implementing it, finding and fixing bugs, and testing. Pandaily reported that the preview sits on a separate branch of the repository and that the current stable 0.27 line continues alongside it. The project is licensed under Apache-2.0.

Alibaba promises scalability and lower costs

Alibaba said version 1.0 is built on an event bus, through which the program’s components pass events to each other following a shared event model. The bus initially runs in the memory of a single process, but the same event model is being extended for distributed use. The aim is to move from one process hosting several sessions to multiple nodes hosting many more, while keeping the existing entry points and return paths.

Alibaba has split task intake, scheduling and execution into separate layers. The company calls the principle separating the brain from the hand: reasoning orchestration and tool execution are managed independently.

The company promises four practical effects:

• Components that handle reasoning, memory and tools can be attached to events, so they interfere less with the main execution path.
• Model calls go through a single entry point that brings together configuration, authorisation and call accounting.
• Sandboxes, or isolated execution environments, are allocated and reused when configuration and permissions are the same. Alibaba said this should reduce repeated initialisation and idle time and so lower costs.
• MCP (Model Context Protocol) connects tools to the agent. Its configuration, authorisation and connections are managed in one place.

The claims about performance and costs are Alibaba’s own. Pandaily reported that the company has not published benchmark results, and the material Pandaily reviewed contained no independent tests. Nor does Alibaba say how closely the open-sourced core matches the version running in Qoder.

The preview is available on GitHub

According to the release notes, Qwen Code 1.0 works with Qwen and Qoder accounts, with several model providers and through users’ own compatible APIs. Users can resume an interrupted session and run the program non-interactively in scripts and automation.

Prebuilt packages for macOS, Linux and Windows do not require a separate Node.js installation and can be downloaded from the Assets section of the release page. The preview’s interfaces, settings and behaviour may still change, and Alibaba has not given a release date for a stable 1.0.

What this is based on

1. github.com

   Publisher: github.com · published 11 November 2026 at 05:55 (UTC+02:00) · retrieved 12 November 2026 at 05:39 (UTC+02:00) · discovery channel: GitHub

2. Alibaba Folds Qoder's Harness Into Qwen Code 1.0 Preview and Open-Sources the Combined Agent Core

   Publisher: pandaily.com · published 12 November 2026 at 05:14 (UTC+02:00) · retrieved 12 November 2026 at 05:38 (UTC+02:00) · discovery channel: web

3. Reportedly: Alibaba Folds Qoder's Harness Into Qwen Code 1.0 Preview and Open-Sources the Combined Agent Core

   Publisher: aigc.news · published 11 November 2026 at 10:11 (UTC+02:00) · retrieved 12 November 2026 at 05:39 (UTC+02:00) · discovery channel: web

DeepSeek published release notes on GitHub on 11 November for a new release candidate of its agent program. The release is version 0.2.2-rc.2 of DeepSeek Harness (DSH), which precedes a stable release.

DSH is an open-source harness: a program that controls an agent, giving the model tools and running its steps.

Subagents share work across several levels

According to the release notes, subagents – separate agents launched for subtasks – can now be named and nested across several levels. Users can enable AgentTeam teamwork per session, and the team has a shared member and task panel.

The change requires migration. AgentTeam now uses unified named subagents, and the interrupt tool is disabled by default. Custom plugins and SDK integrations must move to the new structure, and all processes that share session storage must be upgraded together.

Plugins can also add and remove MCP servers, the services that provide external tools. Plugins can check a server’s status and restart it, so users do not have to maintain each server’s configuration themselves.

Archiving cleans up worktrees by default

Users can now review changes in Git worktrees – parallel working directories of the same repository – by branch, commit and file type. They can also carry out GitHub reviews and update branches directly in DSH.

When a user archives a session, DSH now removes eligible worktrees by default. Branches and commits are preserved. The local directory is kept if cleanup is turned off. Custom configuration and Forge plugins must be migrated.

Changes to search and security

Web search uses the DeepSeek Search API by default. Anyone using older search providers or custom settings must migrate by following the upgrade guide.

The agent runs its commands in a terminal in the sidebar, and the user can take over by typing. If the agent is allowed to read the user’s terminal, the terminal content goes to the model and into session records.

DSH now parses Office documents in its host process, outside the shell sandbox. Old LiveShare permissions must be reconfigured. The PTC feature, which runs code, can access files outside the shared directory, the release notes say, so sessions should be shared only with trusted people.

On Windows, permissions for additional directories now apply only to file tools, not to the terminal or PTC. Custom Computer Use scripts, old Office packages and skills, plugins that rely on the old error classification and Stagehand’s extensionId setting, which no longer works, also need updating. The version also includes, among other things, support for custom model providers that use the Gemini API, as well as bug fixes.

DSH can be installed with npm and used as a desktop app, in the browser, on the command line and through the SDK. In the desktop app’s settings, users can now choose stable or development versions as the update channel. Instructions for a third-party installation on Hugging Face say that using the model requires an API key from DeepSeek or another provider.

What this is based on

1. github.com

   Publisher: github.com · published 11 November 2026 at 13:33 (UTC+02:00) · retrieved 12 November 2026 at 05:42 (UTC+02:00) · discovery channel: GitHub

2. DeepSeek Harness

   Publisher: huggingface.co · source publication time unknown · retrieved 12 November 2026 at 05:42 (UTC+02:00) · discovery channel: web

Aqib Zakaria and Quinn Ennis published their investigation on ChinaTalk on 11 November. They wrote that companies such as Mercor, Surge AI, AfterQuery and Turing are selling roughly half a billion dollars’ worth of data to Chinese labs, and that part of it takes the form of reinforcement learning environments.

The Decoder’s headline on its summary presents the whole sum as environments, although ChinaTalk’s text says environments are only part of the trade.

The same environments as US labs

A reinforcement learning environment is a simulated workplace in which a model attempts a task repeatedly. Grading criteria decide whether the result is rewarded. In ChinaTalk’s example, a vendor builds a PowerPoint clone and hires experts to write criteria for a good slide deck. In a coding environment, the model receives code with an unfixed bug, and the environment checks whether the model finds the fix.

Chinese labs have been reported to buy exactly the same products as Anthropic and OpenAI, the authors wrote, so that they get data proven to work and do not waste compute. According to one study, rollouts take up more than 70 per cent of reinforcement learning time. The savings vary widely from one environment to another, the authors said. Rollouts are memory-bound inference, and Chinese chips are especially short of fast HBM memory.

US labs may pay four or five times the price for exclusivity. Even so, the experience a vendor gains benefits its next customers, the authors wrote.

Nine vendors, partly from anonymous sources

The authors found at least nine vendors: those above plus Abundant, Pareto, DeccanAI, Eaglepoint and Invisible. Named buyers include Alibaba, Tencent, ByteDance, Z.ai and Moonshot AI. The authors do not publish their sources. The evidence rests partly on anonymous sources and on contractors’ LinkedIn posts and CVs. Forbes reported that AfterQuery earns half of its $100 million in annual recurring revenue from China. ChinaTalk, however, could not verify the company’s sales.

The environments are largely outside export controls unless the buyer is on the US Commerce Department’s Entity List, the authors wrote. Buyers can easily sidestep the restriction through a subsidiary outside China that is not on the list. Deals also take place outside China; ByteDance, for example, buys data in Singapore and San Jose. Such deals fall outside China’s procurement registers and most US rules.

The risk lies in cyber and bio

The authors said coding environments can also improve cyber-offensive capabilities, and biology environments can advance gene synthesis and the design of biological weapons. They propose that the Commerce Department’s BIS restrict exports of dual-use environments. Their top priority is for the government to get a better picture of the trade.

The authors do not regard the situation as a catastrophe. Industry sources say Chinese buyers are almost exclusively chasing better benchmark scores, although this is disputed in the industry. Many commercially important benchmarks measure coding and office tasks, so the authors consider the trade likely harmless for now. Buyers also lag by at least a few months, because they get an environment only after a US lab has bought or built it. Few dangerous environments exist, and leading US labs increasingly build environments for sensitive fields themselves.

What this is based on

1. Piggybacking to the Frontier

   Publisher: www.chinatalk.media · published 11 November 2026 at 13:07 (UTC+02:00) · retrieved 12 November 2026 at 05:45 (UTC+02:00) · discovery channel: web

2. US data brokers are selling half a billion dollars in RL training environments to Chinese AI labs

   Publisher: the-decoder.com · published 11 November 2026 at 15:51 (UTC+02:00) · retrieved 12 November 2026 at 05:45 (UTC+02:00) · discovery channel: web

Baidu introduced its new language model on its release page on 11 November. ERNIE 5.5 Flash Preview is the first post-trained version of the ERNIE 5.5 Flash Base model, whose weights Baidu released earlier in November, the company said. The base model only continues the text it is given. Technology site Pandaily reported that the Preview model adds reasoning and tool calls.

Baidu said the model is strong at coding, 3D modelling, knowledge work and content creation. Chinese technology outlet Leiphone reported that the model achieved competitive scores on benchmarks including Terminal-Bench and AutomationBench. Pandaily noted that these are Baidu’s own claims.

Baidu itself warned that the Preview model is an intermediate stage of unfinished post-training and that some of its capabilities may still fall short. The company plans to improve the model based on user feedback before the official release.

Architecture and Baidu’s examples

Baidu’s vLLM-ERNIE repository describes the model as an MoE model: its layers contain 512 expert networks, and ten are selected for each token. The model has 280 billion parameters, but uses 14 billion of them to process each token.

In 37 of the model’s 43 layers, the cache stores the 128 most recent tokens and one compressed entry for each 128-token block. The base model was pretrained with a context of up to 128,000 tokens, but supports a context of about 1 million tokens (1,048,576) at inference. Leiphone reported that the model generates more than 200 tokens per second, but did not give the measurement conditions.

As an example, Baidu said the model wrote a third-person game for Unreal Engine 5.8 in C++ during a single 11-hour coding task. The model worked in a harness, a program that controls the agent, gives the model its tools and runs its work steps. The model handles only text and cannot see the screen. It therefore turned bugs spotted in playtesting, such as a character hovering above the ground, into measurable tests in code.

Availability, pricing and weights

The model is available through the API of Baidu’s Qianfan cloud service. Pandaily reported that Qianfan’s model list, updated on 10 November, gives the ernie-5.5-flash-preview endpoint a default rate limit of 60 requests and 1 million tokens per minute. Leiphone said 1 million input tokens cost 2 yuan at peak times and 1 yuan off-peak. The corresponding output prices are 8 and 4 yuan. A time-limited 50 per cent discount also applies.

Baidu promises to release the weights soon under the Apache 2.0 licence, on 25 November, according to Leiphone. When Pandaily checked on 11 November, the Preview model’s Hugging Face repository returned an access error, and Baidu’s public listing showed only the two versions of the base model.

To run the weights, Baidu offers the vLLM-ERNIE plugin, which requires vLLM 0.32.0 and NVIDIA GPUs. The instructions are in the baidu/vLLM-ERNIE repository on GitHub. They state that the base model’s FP8 weights take up about 280 gigabytes. Running them requires at least two 288-gigabyte, four 192-gigabyte or eight 80-gigabyte GPUs. The 1 million-token context does not fit on 80-gigabyte GPUs.

What this is based on

1. Introducing ERNIE 5.5 Flash Preview · ERNIE

   Publisher: ernie.baidu.com · published 11 November 2026 · retrieved 12 November 2026 at 05:46 (UTC+02:00) · discovery channel: web

2. 百度文心 研究 (ERNIE)

   百度文心 研究 (ERNIE) · Publisher: ernie.baidu.com · source publication time unknown · retrieved 12 November 2026 at 04:01 (UTC+02:00) · discovery channel: web

3. github.com

   Publisher: github.com · source publication time unknown · retrieved 12 November 2026 at 05:47 (UTC+02:00) · discovery channel: GitHub

4. Baidu Releases ERNIE 5.5 Flash Preview on Qianfan, With Open Weights Promised by November 25

   Publisher: pandaily.com · published 11 November 2026 at 09:41 (UTC+02:00) · retrieved 12 November 2026 at 05:47 (UTC+02:00) · discovery channel: web

5. 文心 5.5 Flash Preview发布！编程、3D 建模、知识工作、内容创作等全面升级 | 雷峰网

   Publisher: www.leiphone.com · published 11 November 2026 at 07:22 (UTC+02:00) · retrieved 12 November 2026 at 05:47 (UTC+02:00) · discovery channel: web

Google DeepMind released AlphaGenome 2 on its blog on 11 November. Its predecessor, AlphaGenome, released last year, predicted the effects of changes to DNA letters, or variants, across whole tissues. If a variant acts in only one cell type, the tissue average can mask the effect.

The condition steers the whole model

AlphaGenome 2 reads a DNA sequence of 1 million bases and a single condition: a tissue, cell type or cell state. DeepMind said the condition steers every stage of the model. This lets the model learn how the same DNA is read in different cells.

Researchers can pick the condition from a ready-made list or build it from their own single-cell RNA sequencing, a method that measures gene activity in individual cells. A sample of a few dozen cells is enough, and the model does not need retraining, DeepMind said.

The model predicts gene expression and chromatin accessibility, among other things, and now also DNA methylation. From a single measurement, it can estimate properties that have not been measured. The training data comprised human and mouse tissue and single-cell data: more than 100 million cell profiles from 68 tissues, according to DeepMind.

More hits, but only a minority of causal variants

DeepMind tested the model on the TenK10K dataset, which covers 28 immune cell types. At 90 per cent precision, the model found 30 per cent of the causal variants estimated to alter gene expression. AlphaGenome 1 fine-tuned on single-cell data found 16 per cent. When 90 per cent of direction predictions were correct, the new model identified the direction of effect for 43 per cent of variants, against 23 per cent for its predecessor.

MarkTechPost reported that the model found 41 per cent of variants in a Parkinson’s disease dataset it had not seen before. Scooby, a model fitted to that specific dataset, found 17 per cent. The publication does not say at what precision threshold the figures were calculated.

MarkTechPost notes that the comparisons are DeepMind’s own. Unite.AI reported that the research report warns that even at high precision, the model finds only a minority of statistically likely causal variants. The model may underestimate effects that arise outside open chromatin, in very distant regulatory regions or only after transcription. Outside open chromatin, direction-of-effect predictions reach an AUROC of only 0.63, MarkTechPost said. A value of 0.5 equals guessing.

Unite.AI also said the report states that the cell-state descriptions are based on observational data, so they capture correlations rather than causal relationships. It is uncertain whether the model predicts, for example, the effects of a gene knockout. Transient, stimulus-activated and disease-related cell states are under-represented in the training data and the precomputed scores, and the model does not distinguish fine-grained neuronal subtypes well.

The Verge reported that the new model will not replace its predecessor any time soon. Žiga Avsec, DeepMind’s head of genomics, told the publication that computing predictions with the new model is more expensive.

Access through an API

AlphaGenome 2 is available for non-commercial research via the AlphaGenome API. Unite.AI reported that, according to the documentation, non-commercial use is free. For commercial research, the model is in limited preview. MarkTechPost said the model weights will be released only once the research is published in its final form.

DeepMind is also sharing precomputed results. They include links between regulatory regions and genes in 640 cell types or states and, according to Unite.AI, scores for more than 6.8 million common or trait- and disease-associated variants in 510 states. The model is intended for research use only and has not been validated for clinical use or diagnostics.

What this is based on

1. AlphaGenome 2: Deciphering how cells read DNA to better understand disease — Google DeepMind

   Publisher: deepmind.google · published 11 November 2026 at 17:00 (UTC+02:00) · retrieved 12 November 2026 at 05:45 (UTC+02:00) · discovery channel: web

2. github.com

   Publisher: github.com · source publication time unknown · retrieved 12 November 2026 at 05:46 (UTC+02:00) · discovery channel: GitHub

3. Google’s AlphaGenome 2 zooms in on the cellular roots of disease

   Publisher: www.theverge.com · published 11 November 2026 at 17:00 (UTC+02:00) · retrieved 12 November 2026 at 05:45 (UTC+02:00) · discovery channel: web

4. Google DeepMind Releases AlphaGenome 2 for Cell-Type-Specific Variant Effects

   Publisher: www.marktechpost.com · published 11 November 2026 at 22:51 (UTC+02:00) · retrieved 12 November 2026 at 05:45 (UTC+02:00) · discovery channel: web

5. Google DeepMind’s AlphaGenome 2 Predicts Cell-State Variant Effects

   Publisher: www.unite.ai · published 11 November 2026 at 17:28 (UTC+02:00) · retrieved 12 November 2026 at 05:46 (UTC+02:00) · discovery channel: web

6. 구글 딥마인드, 질병 원인 찾는 AI '알파게놈 2' 공개

   Publisher: zdnet.co.kr · published 12 November 2026 at 02:46 (UTC+02:00) · retrieved 12 November 2026 at 05:46 (UTC+02:00) · discovery channel: web

7. Your heart, brain, and immune system share the same DNA sequence. So how can a single-letter change affect disease in some cells, while sparing the rest? AlphaGenome 2 is our new AI model that…

   GoogleDeepMind · Publisher: x.com · published 11 November 2026 at 17:03 (UTC+02:00) · retrieved 12 November 2026 at 04:13 (UTC+02:00) · discovery channel: X

Researchers from KAIST and DeepAuto.ai say in a preprint published on arXiv on 10 November 2026 that an AI agent’s memory can impair its performance. The work has been submitted to ACL Rolling Review for peer review, which is still under way, arXivDaily reported.

An agent’s memory usually holds insights condensed from earlier runs. A run is the sequence of the agent’s actions and observations. The only thing known about a run’s outcome is whether the task succeeded. The model that writes the insight therefore has to guess which action decided the outcome.

The researchers examined tasks that the agent always solved without memory but failed with it. A language-model judge attributed 52 per cent of these failures to a memory item the agent followed. Most often the item was wrong. The classification describes the content of the agent’s reasoning, not a definite cause.

From insight to source run

WikiTrace stores each insight with the unaltered source run and a summary of why the run succeeded or failed. The agent receives all insights at the start of a task. Before acting, the agent can read the linked files with command-line tools and correct an insight that is incomplete, conflicts with another or comes from a failed run.

Results on five benchmarks

The five benchmarks contained eight suites in total, including text-based household and science tasks and API use. All methods used the same agent and the DeepSeek V4 Flash model, and each test was run three times.

The table shows suite averages. The average is not a single metric, because the suites measure different things, such as success rate or score.

Source: Table 1 of the preprint. Figures are three-run averages across eight suites.

Regression is the share of tasks the agent solved without memory but failed with it. WikiTrace’s regression was 4.5 per cent. With Insights Only, the same insights without source runs, it was 6.5 per cent. When the agent was rerun without memory, 9.7 per cent of successes turned into failures through run-to-run variation alone.

The Insights Only comparison also lacked the prompt that directs the agent to check the sources. The advantage comes from both the source material and this prompt, the authors said.

On three text-based benchmarks, the agent opened the source run in 22 per cent of easy tasks and 44 per cent of hard ones. An easy task was one the agent solved without memory on all three attempts; a hard task was one it never solved.

In openings the agent made after getting stuck or doubting an insight, WikiTrace led Insights Only by 10.4 points, according to the language-model judge’s classification. In other openings the gap was 3.7 points. The agent decided for itself when to open a run, so the gaps do not show that opening caused them.

A link does not prove an insight correct

The link lets the agent check an insight but does not guarantee that the insight is correct, the authors said. The summary’s explanation of a failure may also be wrong. All insights go into the agent’s first message, so the prompt grows with the memory. The authors leave open how the agent would find the insights it needs in a larger memory.

The code is published on GitHub (github.com/yukyeongleee/wikitrace). Installation requires Python, Node, Docker and the uv tool, Blockchain.news reported.

What this is based on

1. Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents

   Lee, Yukyeong, Choi, Yumin, Baek, Jinheon, Hwang, Sung Ju · Publisher: arxiv.org · published 10 November 2026 at 18:46 (UTC+02:00) · retrieved 12 November 2026 at 05:53 (UTC+02:00) · discovery channel: web

2. Paper page - Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents

   Publisher: huggingface.co · source publication time unknown · retrieved 12 November 2026 at 05:55 (UTC+02:00) · discovery channel: web

3. 不要只相信你的记忆：证据关联记忆可降低LLM智能体中的记忆诱导退化｜arXiv 2611.14320｜arXivDaily

   Publisher: www.arxivdaily.com · published 11 November 2026 · retrieved 12 November 2026 at 05:54 (UTC+02:00) · discovery channel: web

4. WikiTrace: Evidence-Linked Memory System Enhances LLM Agent Performance

   Publisher: blockchain.news · published 11 November 2026 · retrieved 12 November 2026 at 05:55 (UTC+02:00) · discovery channel: web

Zihan Dong (Lenovo Group), Qishi Zhan (Marquette University) and Minxuan Hu (Cornell University) published their study on arXiv on 10 November. It is a preprint and has not been peer reviewed.

The researchers re-scored published runs of eight benchmarks. Most agent benchmarks check a run against a task-specific answer key, such as tests derived from a reference solution, and record a single bit: pass or fail. The authors said one bit is not enough to separate the leading agents and gives no extra credit for beating the reference solution.

Gaps are smaller than the errors in the tests

On SWE-bench Verified, the top ten agents’ results fall within 3.6 percentage points, and the largest gap between neighbouring agents is 1.4 percentage points. The ten form a single tier: a group the benchmark cannot statistically separate from the leader.

Stricter SWE-ABS tests rejected 17.7 per cent of the median agent’s accepted fixes. Among the top 30 agents, the stricter tests reversed the order in eight of the 18 neighbouring pairs with a non-zero gap. Removing the same share of passes at random reversed about as many. The order of neighbouring agents is therefore noise, the authors said. The analysis assumes the stricter tests are error-free, and the researchers assessed false rejections only through simulation.

Beating the reference goes unnoticed

In ALE-Bench’s 40 optimisation contests, the official contest score splits the top ten into three tiers. When the same runs were scored as a pass or fail against the median human, all ten fell into one tier, even though 47.8 per cent of runs beat the median human. When the researchers’ graded scoring gave extra credit for exceeding the reference, the top ten split into three tiers again.

Raising the threshold on the same ALE-Bench tasks did not solve the problem. Using the contest winner’s result as the threshold split the top ten of models released after 2025 into two tiers, but 89 per cent of earlier models scored zero on every task. The researchers estimated some of the later models’ release dates retrospectively. They did not simulate a new version that would also add new tasks.

Continuous scoring does not always help: on MLE-bench the answer key alone already separates the top, and on METR’s tasks few runs beat the reference.

The researchers re-ran 13 Terminal-Bench 2.0 tasks with four models, three times each. Three models tied on passes, but on the portfolio-optimisation task the median speedup of claude-opus-5’s solutions was 10.9-fold, against 1.3-fold for the other two. Because there were only three repetitions, the researchers consider the overall separation of the models borderline.

Proposal: acceptance, quality and cost reported separately

The researchers propose first accepting a run with a stricter check. A measured quantity, such as running time, is then scored as a log ratio: the baseline scores zero, the reference solution one, and the cap is three. The baseline and reference are locked and dated so that versions of the benchmark remain comparable.

Cost is reported separately. Four priced agents in the Terminal-Bench 2.0 top ten tied on passes, but their median cost ranged from $0.44 to $2.47 per run.

The scoring method suits only tasks with a measurable quantity. Of Terminal-Bench 2.0’s 89 tasks, 18 compare a number against a threshold. Small samples from SWE-bench Verified, OSWorld-Verified and τ²-bench contained no such tasks, but because the samples were small, the authors cannot rule out a share as high as 10–15 per cent.

What this is based on

1. Agent Benchmarks Stop Ranking Before They Run Out of Tasks

   Dong, Zihan, Zhan, Qishi, Hu, Minxuan · Publisher: arxiv.org · published 10 November 2026 at 06:03 (UTC+02:00) · retrieved 12 November 2026 at 05:55 (UTC+02:00) · discovery channel: web

2. Agent Benchmarks Stop Ranking Before They Run Out of Tasks

   Publisher: www.alphaxiv.org · published 10 November 2026 · retrieved 12 November 2026 at 05:55 (UTC+02:00) · discovery channel: web

Researchers from Zhejiang Gongshang University, the University of Sussex and two other organisations published a preprint on arXiv on 9 November that has not been peer reviewed. It measures how deleting reasoning traces from earlier turns changes how often a reasoning model succeeds.

A reasoning trace is the text a model writes before it acts. In a multi-turn session, the model reads the earlier traces on the next turn. Providers tie a trace to whoever produced it, the authors wrote, so a request routed to another provider arrives without the traces.

One session, three ways

The researchers ran 794 multi-turn tool tasks from the BFCL v4 benchmark on seven open-weight models. They ran each task three times: keeping the traces, deleting the traces from all earlier turns, and deleting traces only from the early part of the session, about 40 per cent of turns. Tool calls with their results and the trace of the current turn were always kept. BFCL’s official checker scored the session’s final state without knowing how the traces had been handled. The researchers repeated the runs with three random seeds. Baichuan-M2-32B and Olmo-3-7B-Think were run on only 200 tasks.

The effect varied by model

The table compares success rates when all earlier traces were kept or deleted.

Source: Jia et al., arXiv 2611.13223. Pooled results across three seeds. A positive difference means deletion lowered the score.

R1-Distill-14B improved from a baseline at which it succeeded on 1.3 per cent of tasks. Olmo’s difference rests on ten and two successes. Model size did not explain the differences: the three 32-billion-parameter models differed by −0.2, +0.7 and +51.3 percentage points.

With Qwen3-32B, the researchers deleted 0–100 per cent of the history on 200 tasks. Success rates at the two extremes differed by 1.5 percentage points. The models sensitive to deletion were not tested this way.

Repeated calls and silent failures

Deletion also changed how the models behaved. After early-session traces were deleted, R1-Distill-14B repeated the same call with the same arguments in 68.7 per cent of sessions, against 18.1 per cent with traces kept. For Olmo, deleting all traces raised the share from 30.2 to 66.5 per cent. Baichuan repeated itself only slightly more, yet still failed. Qwen3-8B made more calls after deletion and succeeded just as often. Whether a model re-derives the deleted reasoning remains open, the authors said.

The authors interpret sensitivity as a property of the model and propose a screen: run the model on 200 tasks with traces kept and deleted. The history is kept if the confidence interval of the difference lies entirely above 10 percentage points. That requires a point estimate above roughly 20 percentage points. The screen would therefore likely class a model whose score drops by, say, 15 percentage points as tolerant of deletion.

The results apply to BFCL-style function-calling tasks. Whether they hold for other tasks is an open question, the authors said. Closed models were not measured. Session logs are available from the authors on request.

What this is based on

1. Deletion Sensitivity Is a Property of the Model: A Three-Condition Measurement across Seven Open Reasoning Models

   Jia, Hanzhang, Zeng, Liheng, Yang, Zhiyu, Gao, Yi, Ma, Bo · Publisher: arxiv.org · published 9 November 2026 at 22:51 (UTC+02:00) · retrieved 12 November 2026 at 05:55 (UTC+02:00) · discovery channel: web

2. Deletion Sensitivity Is a Property of the Model | OpenTrain AI

   Publisher: www.opentrain.ai · published 9 November 2026 at 22:51 (UTC+02:00) · retrieved 12 November 2026 at 05:55 (UTC+02:00) · discovery channel: web

3. Deletion Sensitivity Is a Property of the Model: A Three-Condition Measurement across Seven Open Reasoning Models

   Publisher: www.alphaxiv.org · published 9 November 2026 · retrieved 12 November 2026 at 05:55 (UTC+02:00) · discovery channel: web

Tom Aarsen described the method on the Hugging Face blog on 11 November and released training and evaluation scripts along with two example models. Aarsen says the work builds on Stefan Josef’s Learn to Pool and on research by LightOn. Aarsen’s own aim is to explain the method and provide a reproducible example.

A multi-vector model such as ColBERT stores a separate vector for every token in a document. At search time, each query token picks the best-matching document vector, and the scores are summed. The vectors add up: in Aarsen’s calculation, the 4-bit quantised residuals of 128-dimensional vectors take up 12.8 GB for a collection of 1 million documents with 200 vectors per document. With 50 vectors, 3.2 GB is enough.

How pooling and training work

Users can shrink the index by merging similar vectors within a document (token pooling). HierarchicalTokenPooling in Sentence Transformers groups them with Ward’s hierarchical clustering and replaces each group with its mean. The number of vectors can be limited by a ratio (pool_factor), a maximum or a distance threshold. The threshold can be calibrated to produce a target average number of vectors. Because calibration targets the average, the count varies from document to document, and on another dataset even the average may change. Queries are not pooled by default.

A standard model is not trained to work with pooled vectors, so merging can erase a distinction that matters for retrieval. In pooling-aware training, the loss function MultiVectorTokenPoolingLoss computes the retrieval loss at selected pooling levels. Aarsen’s example used two levels, unpooled and pooled to about a quarter, weighted equally. The group assignment is not differentiated; instead, gradients pass from the pooled vector straight to the group’s members (straight-through estimator). The pooled versions are built from vectors already computed, so the model does not need a separate pass for them.

Results from a single comparison run

Aarsen trained two models from ModernBERT-base with the same settings on the same 120,000 MS MARCO examples for one epoch. Only the loss function differed, and no hyperparameters were tuned. The metric was mean nDCG@10 across NanoBEIR’s 13 datasets.

In the table, the pooling-aware model scored higher at every pooling level tested, with the largest gap at the tightest limit.

Source: Aarsen, Hugging Face blog. NanoBEIR, 13 datasets, mean nDCG@10 (%).

The table cannot show whether the advantage holds on other datasets, because the models were tested only on NanoBEIR. Each model had a single training run, and the figures are the author’s own.

Pooled with a factor of 4, the pooling-aware model retained 97.8 per cent of its own unpooled score and the standard model 94.6 per cent. The nDCG@10 scores then differed by 2.2 points (49.03 and 51.23), with 74.8 per cent fewer vectors than without pooling. The advantage also appeared with maximum limits and calibrated thresholds, even though training used only ratio-based pooling.

Outlets including zglg.work and AI Breaking Wire repeat the blog’s figures. Whether the advantage holds in much larger datasets and production indexes has yet to be established, zglg.work wrote.

The method is part of the Sentence Transformers library, and pooling can also be used with conventionally trained multi-vector models. The blog advises running the training script on a local GPU or one from Hugging Face Jobs. For optional GPU acceleration, the library picks compatible GPU kernels and uses the standard implementation elsewhere; the default MaxSim scoring kernel, for example, requires CUDA and Triton.

What this is based on

1. Token Pooling-Aware Training for Multi-Vector Embedding Models

   Publisher: huggingface.co · published 11 November 2026 at 02:00 (UTC+02:00) · retrieved 12 November 2026 at 05:56 (UTC+02:00) · discovery channel: web

2. Hugging Face ships token pooling-aware training for multi-vector retrieval models

   Publisher: zglg.work · published 11 November 2026 at 02:00 (UTC+02:00) · retrieved 12 November 2026 at 05:57 (UTC+02:00) · discovery channel: web

3. Token Pooling-Aware Training for Multi-Vector Models

   Publisher: aibreakingwire.com · published 11 November 2026 at 18:30 (UTC+02:00) · retrieved 12 November 2026 at 05:57 (UTC+02:00) · discovery channel: web

4. Token Pooling-Aware Training for Multi-Vector Embedding Models — GazeWire

   Publisher: gazewire.com · published 11 November 2026 at 17:26 (UTC+02:00) · retrieved 12 November 2026 at 05:57 (UTC+02:00) · discovery channel: web

DualEntry announced the results on 11 November in a paid Business Wire press release, which Yahoo Finance also published. The benchmark is the company’s own, not an independent measurement. DualEntry sells mid-sized companies an ERP system in which AI agents do accounting work.

Google’s new Gemini 4 Carbon scored 88.1 per cent and Gemini 4 Argon stayed on top with 91.1 per cent, the release said. Carbon shares second place with Zhipu AI’s GLM-5.3 Flash and SpaceXAI’s Grok 4.7.

101 invented accounting cases

The benchmark contains 101 synthetic accounting cases, written for the test, across eight areas. These include journal entries, accounts payable and receivable, bank reconciliation and month-end close. No customer data is used.

Each model runs the whole benchmark three times with the same test company, tools and instructions. Fixed scoring rules compare the answers with reference answers written in advance. No AI model acts as a grader.

The table shows overall scores for the models named in the release. Argon leads the three models in joint second place by three percentage points, so a single benchmark from one company cannot show which models are better at accounting work in general.

Source: DualEntry press release, 11 November 2026. Of the 24 models tested, the table lists those named in the release.

Carbon solved 25 of 28 hard tasks, or 89.3 per cent, the best hard-task result in the benchmark. Argon solved 82.1 per cent of them, two tasks fewer. The release said Carbon did slightly better on hard tasks than on easy ones, which is unusual.

Carbon solved every transaction classification task. Its weakest area was bank reconciliation at 66.7 per cent, where Argon scored 75 per cent and Moonshot AI’s Kimi K3.1 Pro 100 per cent.

Carbon cheaper, Argon more consistent

Argon had the benchmark’s best consistency score, 86.1 per cent; Carbon scored 80.2 per cent. Benedict Dohmen, a co-founder of DualEntry, said Argon gave the same answer in all three runs on 86 per cent of tasks. Carbon costs $0.081 per task, about 30 per cent less than Argon’s $0.115.

In the release, Dohmen said consistency is the next decisive difference for finance teams, and that an AI agent should match Argon’s consistency before it gets access to the books.

DualEntry’s page dualentry.com/benchmark lists results for all 24 models, including prices and response times.

What this is based on

1. Google Now Has Two of the Top AI Models for Accounting Work, DualEntry Benchmark Finds

   Publisher: uk.finance.yahoo.com · published 11 November 2026 at 15:17 (UTC+02:00) · retrieved 12 November 2026 at 05:47 (UTC+02:00) · discovery channel: web

Intuit announced the Agentic Books Close pilot on 9 November in a post on its Tax Pro Center blog presenting the autumn release of Intuit Accountant Suite. Intuit Accountant Suite is software for accounting firms. Intuit says it forms a unified platform with clients’ QuickBooks accounting software.

The month-end close finalises the month’s books: posting transactions, reconciling accounts and preparing reports. Intuit said the agent handles the preliminary preparation in the background, in line with the firm’s own standards. The accountant is left with review, approval and exception management.

As Intuit describes it, the agent posts pending transactions, reconciles accounts, flags anomalies and recommends fixes. It also identifies fixed assets, prepaids and other accruals. For these items, it drafts schedules and journal entries, which are posted only after the accountant approves them.

The agent also assembles a monthly package that firms can send directly to clients, containing Variance Analysis and AP/AR Aging reports.

Every AI-drafted entry must be reviewed and approved before posting, and the agent’s actions are recorded in an audit log, the company said. Intuit itself warns that the agent is not always accurate. According to the company, the agent learns from the accountant’s edits with each close.

The agent is a separate product from Books Close, which is now in beta in the Intuit Accountant Suite Core and Accelerate tiers at no additional cost. Books Close features include close review templates, task tracking, AI-based anomaly detection and automatic updates to the client’s QuickBooks file.

The autumn release includes other beta features too: queries across the whole client base on the Intuit Intelligence AI platform, and Playbooks templates that let firms standardise onboarding for new clients. These are part of the Accelerate tier, which is free until January 2027.

The pilot is open to a limited number of accounting firms, which can apply on the Intuit Accountant Suite feature preview page. The agent is free during the pilot. After the pilot and beta, firms will need a paid add-on to keep using it; Intuit has not disclosed the price. The company said features and release dates may still change.

What this is based on

1. What’s new in Intuit Accountant Suite - Tax Pro Center | Intuit

   Publisher: accountants.intuit.com · published 9 November 2026 at 16:00 (UTC+02:00) · retrieved 11 November 2026 at 05:50 (UTC+02:00) · discovery channel: web

Copenhagen-founded Sumary announced its funding round on 10 November 2026. The investment firm Singular led the round, and the company’s existing investors byFounders, Partech and Tenity also took part.

The amount was announced in dollars. Sesamers converted it to about €22.3 million, and the British publication Soapbox puts it at £19 million. The company says it is the largest seed round for an accounting software company in Europe. Sesamers repeats this claim, citing Oresund Startup News.

The round comes less than a year after Sumary’s previous funding: in November 2025 it raised a $4.2 million pre-seed round, which byFounders led, The Next Web reported.

Sumary’s round was the third European accounting funding round within 12 days, Sesamers reported. Before that, Belgium’s Ravical raised $17 million and Barcelona-based Kabilio raised €10 million.

An accountant approves every step

Sumary calls its product an operating system for accounting firms. The platform automates routine work: the monthly close, in which the month’s entries are reconciled and closed; VAT returns; and preparation of annual accounts. It works with clients’ existing accounting software and prepares tasks for an accountant to review and approve.

The company’s research team in Copenhagen has built a benchmark from real accounting tasks. Sumary uses it to measure leading AI models and tests every new model before it can handle a client’s books. According to company information published by Soapbox and FinTech Global, the benchmark also assesses the financial data given to the model, the controls applied to its work and the checks on its output.

Time savings are the company’s own figures

Sumary says that after three months of use, accounting firms spend 50 per cent less time on the monthly close and 75 per cent less on VAT returns. Sesamers notes that the figures have not been independently verified. The published information does not say what baseline the savings are measured against or how many firms’ experience they are based on.

The company says its clients include most of the 20 largest accounting firms in the Nordic countries, as well as Deloitte, Grant Thornton and Baker Tilly, among others.

AI-based accounting firms challenge established ones

Chief executive Pascar Sivam points to labour shortages as the reason for the product: ‘Accounting firms are being asked to do more with fewer people.’ Software must handle the preparatory work properly so that accountants have time for their clients and their own teams, Sivam says.

‘Dozens, and probably soon hundreds, of AI accounting firms are being built as we speak,’ Sivam says. Sumary wants to help established firms adopt AI quickly across their whole operations and so hold their own against these newcomers.

Sumary will use the funding to expand in Europe, starting with the Benelux countries, and to grow sales in the UK. London will become its commercial hub, and Sivam and chief operating officer Frederik Skov Wacher are moving there. Technology development and AI research will stay in Copenhagen.

What this is based on

1. Sumary raises $25M seed to scale AI accounting in Europe

   Publisher: www.sesamers.com · published 11 November 2026 at 09:08 (UTC+02:00) · retrieved 12 November 2026 at 05:52 (UTC+02:00) · discovery channel: web

2. Sumary secures $25M to expand its accounting software across Europe

   Publisher: tech.eu · published 10 November 2026 at 09:21 (UTC+02:00) · retrieved 12 November 2026 at 05:52 (UTC+02:00) · discovery channel: web

3. Sumary raises $25M in what it calls Europe's largest accounting seed round

   Publisher: thenextweb.com · published 10 November 2026 at 09:45 (UTC+02:00) · retrieved 12 November 2026 at 05:52 (UTC+02:00) · discovery channel: web

4. Sumary raises a £19m seed round led by Singular to run core workflows for accounting firms

   Publisher: www.soapbox.vc · published 10 November 2026 · retrieved 12 November 2026 at 05:52 (UTC+02:00) · discovery channel: web

5. Sumary raises $25m to challenge Europe’s accounting sector

   Publisher: fintech.global · published 10 November 2026 at 14:53 (UTC+02:00) · retrieved 12 November 2026 at 05:52 (UTC+02:00) · discovery channel: web

New Zealand-based LawVu announced the new integration in a press release on 11 November 2026. The tools are Anthropic’s Claude, Microsoft Copilot, OpenAI’s ChatGPT and the legal AI services Harvey and Legora. The connections are available immediately and require no custom development, the company said.

LegalOS is a system built for in-house legal teams. It stores legal matters, contracts, legal spend and the company’s own legal knowledge, among other things. Unite.AI reported the news the same day, based on the press release and LawVu’s documentation. The information therefore comes from the company and has not been independently verified.

What the tools can do

The connection runs through LawVu’s MCP server. MCP (Model Context Protocol) is an open standard for connecting AI tools to external systems. The server offers more than 50 actions for reading and updating LegalOS data, LawVu said.

Based on LawVu’s documentation, a tool can, for example, search and summarise legal matters and contracts and read key contract dates, Unite.AI reported. It can also create matters and contracts, log status updates, assign tasks and start contract workflows. In LawVu’s own example, a lawyer asks Copilot which vendor contracts auto-renew next quarter. The lawyer then assigns a review task to a colleague without leaving Copilot.

Permissions stay at user level, but data goes to the AI provider

The documentation promises that an AI tool gets no broader permissions than its user has, Unite.AI reported. A user who cannot open a contract in LawVu cannot reach it through the AI tool either. Actions taken through MCP appear in LawVu’s activity log in the same way as actions performed directly on the platform.

LawVu’s control does not, however, extend to everything that happens to data once it leaves the platform. Data involved in a query is passed to the selected AI provider for processing. LawVu does not control what the provider does with it. The company advises customers to check the provider’s processing terms and privacy practices against their own obligations.

The documentation lists prompt injection, meaning instructions hidden in content that can steer the model, as a risk. Other risks include third-party MCP servers, misleadingly named tools and unwanted autonomous actions. LawVu recommends giving tools only the permissions they need and using vetted tools and servers. It also recommends monitoring logs, rolling the feature out to a limited group first and, where possible, having a human approve AI changes to records.

The feature is not switched on automatically. Customers must request it from LawVu, and an organisation administrator must enable it in the Integrations view. The AI service must also be configured. Unite.AI reported that Claude and Harvey have official connectors, and ChatGPT and Codex an official plugin. The list does not mention Copilot or Legora. Other MCP-compatible services connect using LawVu’s server URL, with OAuth 2.0 authentication where required.

What this is based on

1. LawVu Gives Five AI Tools Governed Access to LegalOS Workflows

   Publisher: www.unite.ai · published 11 November 2026 at 18:15 (UTC+02:00) · retrieved 12 November 2026 at 05:52 (UTC+02:00) · discovery channel: web

2. LawVu MCP Server connects Claude, Copilot, ChatGPT and more

   Publisher: lawvu.com · published 11 November 2026 at 16:00 (UTC+02:00) · retrieved 12 November 2026 at 05:53 (UTC+02:00) · discovery channel: web

Harvey, which develops AI tools for legal work, and Danish legal publisher Djøf Forlag announced their partnership in press releases on 11 November. The content comes from Jurabibliotek, Djøf Forlag’s online service for Danish legislation.

The content covers consolidated acts, the executive orders (bekendtgørelser) issued under them, amendments and preparatory works. Preparatory works are documents from the legislative process. They show what the legislature intended a provision to achieve, and lawyers return to them when interpreting the law.

Harvey said a lawyer can move from a provision to the executive orders that implement it and to the preparatory works, with each answer citing its source. In the company’s examples, a user checks what the preparatory works said when a provision was enacted, or drafts advice on Danish law based on these sources and the firm’s previous work.

The content can be used in Harvey Assistant, the Agents feature, Vault (Harvey’s document storage and analysis service), Workflows and the Word Add-in. Lawyers can use it alongside their team’s own notes, prior advice and documents. Harvey already includes Danish public legal sources.

Bjarke Madsen, digital director at Djøf Forlag, cited demand as the reason for the deal: customers increasingly want the publisher’s sources in AI-powered legal tech systems. Madsen said the publisher has provided Danish legal sources for more than 65 years.

The press releases also quote Accura, a Danish law firm and Harvey customer. Its head of AI, Christian Høyer Ørum, described how the firm works: ‘When we advise on a regulatory question, we start with the act, check the executive orders, and go back to the preparatory works when interpretation matters.’ In Ørum’s view, doing all of this in one tool makes the work faster and more consistent.

Harvey previously formed a data partnership with Gyldendal Rettsdata, a Norwegian legal information service.

Harvey said Jurabibliotek’s legislative content will be available at the start of next year at no additional charge. Access requires a Jurabibliotek licence.

What this is based on

1. Harvey Adds Danish Legislation and its Preparatory Works Through Partnership with Djøf Forlag

   Publisher: www.harvey.ai · published 11 November 2026 at 09:34 (UTC+02:00) · retrieved 12 November 2026 at 05:58 (UTC+02:00) · discovery channel: web

2. Harvey Partners with Djøf Forlag to Add Danish Legal Sources

   Publisher: www.unite.ai · published 11 November 2026 at 09:52 (UTC+02:00) · retrieved 12 November 2026 at 05:58 (UTC+02:00) · discovery channel: web

3. Djøf Forlag indgår samarbejde med Harvey

   Publisher: djoefforlag.dk · published 11 November 2026 at 11:53 (UTC+02:00) · retrieved 12 November 2026 at 05:59 (UTC+02:00) · discovery channel: web

4. Harvey is adding Djøf Forlag’s Danish legal sources to its platform

   Publisher: airadarhaber.com · published 11 November 2026 at 11:08 (UTC+02:00) · retrieved 12 November 2026 at 05:59 (UTC+02:00) · discovery channel: web

5. Harvey Adds Danish Legislation and its Preparatory Works Through Partnership with Djøf Forlag (Harvey Ai Corp)

   Publisher: article.wn.com · published 11 November 2026 · retrieved 12 November 2026 at 05:59 (UTC+02:00) · discovery channel: web

Anthropic set out its compute strategy in an investor video that is part of its initial public offering roadshow, which began on 10 November. The next day, software developer Simon Willison described the video in a blog post, and The Decoder, an AI news site, reported on it. There is no official transcript, so Willison transcribed the video himself with the speech recognition program MacWhisper.

In the video, Anthropic co-founder Tom Brown said the company trains and runs its Claude model on chips from four manufacturers. The Decoder describes Brown as the company’s head of compute. The chips are AWS’s Trainium, Google’s TPU and GPUs (graphics processing units) from Nvidia and AMD. Each chip type needs its own software, which is expensive and difficult to build, Brown said.

Three days instead of months

Anthropic began using AMD GPUs this year. In the past, Brown said, bringing up a new chip platform took dozens of engineers months. This time the company’s Mythos model did the work, getting Claude running on AMD chips in three days with virtually no human help.

The model also ported or wrote almost all the kernels, Brown said. Kernels are small programs that carry out a model’s calculations on a particular chip.

A chart in the investor presentation compares the three days with 6–12 months, The Decoder reported. The chart gives that as the time traditional engineering needs to bring new hardware to usable performance. Its figures concern prompt processing on AMD MI355X GPUs, so the comparison is limited to one stage of inference on one GPU model.

The Decoder notes that the figures are Anthropic’s own and have not been independently verified. In the publication’s view, if the figures hold, Nvidia’s strongest competitive advantage, the CUDA software ecosystem that ties developers to its chips, would lose importance. Companies could then switch chip suppliers more easily than they can today.

Compute multipliers

Brown named compute multipliers as a second advantage. These are research findings that let a company get more out of the same physical computing capacity. In Brown’s example, a fourfold multiplier means a model that needs 100 units of compute can be trained with 25. The multipliers reinforce one another, Brown said.

Anthropic believes Claude will eventually find multipliers on its own as development approaches recursive self-improvement, Brown said. The Decoder notes that the idea is disputed among researchers and has not been proven.

Both GPU makers have financial ties to Anthropic, The Decoder reported. In November 2025 Nvidia committed to invest up to $10 billion in the company. In July 2026 AMD announced an investment of up to $5 billion. In return, Anthropic plans to deploy up to two gigawatts of AMD MI455X GPUs, with the first gigawatt phase beginning in the first half of 2027.

What this is based on

1. Anthropic’s IPO roadshow video

   Publisher: simonwillison.net · published 11 November 2026 at 17:11 (UTC+02:00) · retrieved 12 November 2026 at 05:28 (UTC+02:00) · discovery channel: web

2. Anthropic says AI is eroding Nvidia's CUDA lead, cutting an AMD chip switch from months to three days

   Publisher: the-decoder.com · published 11 November 2026 at 16:25 (UTC+02:00) · retrieved 12 November 2026 at 05:28 (UTC+02:00) · discovery channel: web

The Wall Street Journal and The Information have reported on OpenAI’s finances for July–September. Heise online said the WSJ’s figures were based on material OpenAI had given its investors. Bloomberg reported the revenue figure, citing anonymous sources. TechCrunch reported the figures with a caveat: if they are true.

Quarterly revenue exceeded $10 billion for the first time, Bloomberg reported. Earlier reports put April–June revenue at $6.7 billion, up 17–18 per cent on the previous quarter. Figures cited by The Decoder put July–September revenue a year earlier at $3.9 billion.

The loss grew but shrank relative to sales

The adjusted operating loss grew from $10.1 billion to $10.9 billion, the WSJ reported. The figure excludes stock-based compensation, which totalled $2.6 billion.

The adjusted operating margin improved from −150 per cent to −109 per cent. A year earlier it was −97 per cent.

Anthropic is still ahead

Anthropic’s revenue grew 47 per cent to $17 billion in the same quarter. Its operating loss was $567 million. The losses are not directly comparable: the WSJ said Anthropic’s figure includes stock-based compensation and OpenAI’s does not.

The annual revenue run rate is the latest period’s sales scaled up to a full year. Bloomberg put OpenAI’s run rate at $50 billion. Anthropic said its run rate was $81 billion at the end of October. The companies calculate the figure differently, TechCrunch reported: Anthropic includes sales made through its cloud partners, while OpenAI counts only its own share of some partner sales.

Brockman: enterprise sales doubled

OpenAI’s president, Greg Brockman, recently said the first two quarters of the year had been difficult, Heise online reported. Brockman nevertheless spoke of an incredible turnaround in revenue. The Decoder reported Brockman as saying that the enterprise business doubled in the third quarter and now accounts for most of the company’s revenue.

According to third-party data cited by the WSJ, companies looked for cheaper alternatives to Anthropic’s Claude models, which boosted OpenAI’s API sales. This year OpenAI has cut prices for its new models, shut down its video app and invested in its coding tool Codex, The Decoder reported.

ChatGPT has about 100 million paying subscribers, or about 8.5 per cent of its weekly users, the WSJ reported. The Decoder put weekly users at 1.2 billion and said ChatGPT advertising had an annual revenue run rate of $1 billion in August.

OpenAI is raising funding at a valuation of $1.25 trillion but has secured less than half of the $30 billion it is seeking, The Decoder reported. Heise online said OpenAI plans to go public in 2027 at the earliest, while The Decoder said Anthropic plans to list as early as next week.

What this is based on

1. Rapidly scaling online storage to serve over 1 billion ChatGPT users (Part II)

   Publisher: openai.com · source publication time unknown · retrieved 12 November 2026 at 05:43 (UTC+02:00) · discovery channel: web

2. Enterprise sales push OpenAI's quarterly revenue to $10 billion, though it still trails Anthropic

   Publisher: the-decoder.com · published 11 November 2026 at 14:12 (UTC+02:00) · retrieved 12 November 2026 at 05:43 (UTC+02:00) · discovery channel: web

3. OpenAI's quarterly revenue reportedly crosses $10B for the first time | TechCrunch

   Publisher: techcrunch.com · published 11 November 2026 at 18:14 (UTC+02:00) · retrieved 12 November 2026 at 05:43 (UTC+02:00) · discovery channel: web

4. OpenAI: 10 Milliarden US-Dollar Umsatz, 11 Milliarden US-Dollar Verlust

   Publisher: www.heise.de · published 11 November 2026 at 08:17 (UTC+02:00) · retrieved 12 November 2026 at 05:43 (UTC+02:00) · discovery channel: web

5. OpenAI Revenue Jumps to $10 Billion in Third Quarter Amid Growing Losses

   Publisher: www.theinformation.com · published 10 November 2026 at 23:07 (UTC+02:00) · retrieved 12 November 2026 at 05:43 (UTC+02:00) · discovery channel: web

Bloomberg reported the oversubscription on Wednesday 11 November, a day after the offering’s investor roadshow began. Bloomberg’s sources, as cited by Kauppalehti, said large preliminary subscription intentions from anchor investors had driven demand for shares in Anthropic, the developer of the Claude AI model. Anchor investors are large investors who state their intention to subscribe early in an offering.

Oversubscription means investors want to buy more shares than are on offer. Bloomberg did not say by how much demand exceeds supply, Gizmodo reported. Bloomberg’s sources noted that details of the offering could still change. Anthropic did not respond to Bloomberg’s request for comment.

Part of the proceeds goes to existing shareholders

The offering comprises 825 million shares priced at $70–78, Kauppalehti reported. At the top of the range, the sale is worth about $64.4 billion before the over-allotment option. That sum includes shares sold by existing shareholders, and Anthropic receives none of those proceeds. The lead banks, Goldman Sachs and Morgan Stanley, are expected to receive an option to buy up to 123.75 million additional shares within 30 days, Gizmodo reported.

Anthropic is also carrying out a $26 billion private placement, selling shares to pre-selected investors. Buyers include Nvidia, AMD, Alphabet, Broadcom, Micron, Microsoft and TSMC, Arvopaperi reported. Together, the IPO and the private placement are worth up to $90.4 billion.

Kauppalehti said the price range corresponds to a market value of $1,260–1,410 billion when employee stock options are included. The paper said the offering would be the largest in stock market history; the previous record was set by SpaceX’s listing in the summer. SpaceX’s market value then rose to $1,770 billion, Gizmodo reported.

Banks consider raising the price

The banks running the listing are considering raising the price range because of strong demand, the Financial Times reported, citing five sources. Arvopaperi and Kauppalehti, citing the FT, reported that the valuation could then rise to $1,500 billion or more. Arvopaperi added that, based on the FT’s information, the amount raised in the IPO could exceed $100 billion. No new range has been published.

On 10 November, the board of Taiwanese chipmaker TSMC approved an investment of up to $500 million in Anthropic shares at the expected IPO price, AI Times reported. TSMC may invest the same amount in OpenAI preferred stock, Ars Technica reported, citing Reuters.

Anthropic and OpenAI buy hardware made by TSMC for their data centres, Ars Technica reported. Citing the Wall Street Journal, Ars Technica said analysts warn that mutual financing between investors, suppliers and customers makes it harder to gauge real demand for AI.

Anthropic is seeking a listing on the Nasdaq Global Select Market under the ticker ANT, CNBC reported. The offering for Australian retail investors opened on Wednesday and closes on 17 November, according to broker Stake’s timetable. Retail investors can also buy shares directly in Japan, Canada and parts of Europe. Anthropic chief executive Dario Amodei and JPMorgan Chase chief executive Jamie Dimon will appear at a livestreamed investor event on Friday, Gizmodo reported.

What this is based on

1. Bloomberg: Anthropicin anti ylimerkitty moninkertaisesti

   Publisher: www.kauppalehti.fi · published 11 November 2026 at 15:34 (UTC+02:00) · retrieved 12 November 2026 at 05:43 (UTC+02:00) · discovery channel: web

2. FT: Anthropicin listautumishinta voi nousta heti

   Publisher: www.arvopaperi.fi · published 11 November 2026 at 08:46 (UTC+02:00) · retrieved 12 November 2026 at 05:43 (UTC+02:00) · discovery channel: web

3. MARKKINAT: Aasian siruosakkeet nousivat Wall Streetin laskusta huolimatta – Arviot Anthropicin arvosta ylittävät jo 1 500 miljardia dollaria

   Publisher: www.kauppalehti.fi · published 11 November 2026 at 07:48 (UTC+02:00) · retrieved 12 November 2026 at 05:43 (UTC+02:00) · discovery channel: web

4. Anthropic opens Australian retail IPO offering as AI scrutiny intensifies

   Publisher: www.cnbc.com · published 11 November 2026 at 03:26 (UTC+02:00) · retrieved 12 November 2026 at 05:44 (UTC+02:00) · discovery channel: web

5. Taiwan’s chipmaking giant TSMC to invest $1B in Anthropic, OpenAI

   Publisher: arstechnica.com · published 11 November 2026 at 20:16 (UTC+02:00) · retrieved 12 November 2026 at 05:44 (UTC+02:00) · discovery channel: web

6. TSMC, 자체 칩 개발 나선 오픈AI·앤트로픽에 최대 1.3조원 투자 - AI타임스

   Publisher: www.aitimes.com · published 11 November 2026 at 10:05 (UTC+02:00) · retrieved 12 November 2026 at 05:44 (UTC+02:00) · discovery channel: web

7. Investors Are Lining Up to Throw Money at Anthropic

   Publisher: gizmodo.com · published 11 November 2026 at 20:00 (UTC+02:00) · retrieved 12 November 2026 at 05:44 (UTC+02:00) · discovery channel: web

8. Anthropic’s IPO Is Said to Be Multiple Times Oversubscribed

   Publisher: www.bloomberg.com · published 11 November 2026 at 15:17 (UTC+02:00) · retrieved 12 November 2026 at 05:43 (UTC+02:00) · discovery channel: web

The European Commission has put forward a bill to overhaul consumer protection, the technology site Ghacks reported on 11 November. The proposal covers online shopping, social media, apps and digital subscriptions.

Consumer Protection Commissioner Michael McGrath called the package ‘the most significant reform of EU consumer protection rules in more than 20 years’, Ghacks reported. McGrath said the current rules were written before the growth of online shopping and of services that compete for users’ attention.

What the proposal would ban or require

Dark patterns are interface designs that steer users into decisions they would not otherwise make. The proposal would ban them. Ghacks gives the example of a false claim that a product is running out.

The proposal would also:

• restrict dynamic pricing by ticket sellers, meaning prices that change with demand
• require influencers to disclose when they are paid to promote a product
• ban personalised advertising and pricing that target minors or use sensitive data
• make digital subscriptions easier to cancel
• give people the right to deal with a human in customer service.

All users could also turn off addictive features such as video autoplay and streaks, which count consecutive days of use. Ghacks said this would affect apps such as Duolingo and Strava. The EU already restricts some of these features for children through the Kids Act, the site said, and the proposal would extend similar choices to adults.

Fines and links to other EU laws

The Commission would gain direct powers to investigate cases with a large-scale impact and to fine companies up to 6 per cent of their worldwide annual turnover. Companies based outside the EU would have to appoint a legal representative in the EU.

Executive Vice President Henna Virkkunen said the act is intended to fill gaps between three existing laws rather than duplicate them: the Digital Services Act (DSA), the Digital Markets Act (DMA) and the AI Act.

The DSA already covers some addictive design on large platforms. Last Friday the Commission found that TikTok’s For You feed had breached the act, Ghacks reported. In July the Commission reached the same finding on similar Meta features on Instagram and Facebook. Each company could face a fine of up to 6 per cent of its worldwide turnover if the findings are confirmed.

Nothing changes for users yet

The Digital Fairness Act is a proposal, not yet law. For now, apps, ticket sellers and subscription services can keep their current designs. If the rules are adopted as they stand, Ghacks said, the clearest changes for users would be a setting to turn off streaks and autoplay, an easier way to cancel subscriptions, and fewer countdown timers and ‘Only 2 left’ notices in online shops.

The Commission has not said when the rules would take effect if adopted.

What this is based on

1. EU Proposes Digital Fairness Act to Ban Dark Patterns and Let Users Turn Off Streaks and Autoplay

   Publisher: www.ghacks.net · published 11 November 2026 at 16:58 (UTC+02:00) · retrieved 12 November 2026 at 05:44 (UTC+02:00) · discovery channel: web

Financial services company Ramp reported on 11 November how AI payments by its business customers changed in October. The Decoder and The Neuron covered the figures. The Ramp AI Index tracks payments by more than 85,000 US business customers, and Ramp’s lead economist Ara Kharazian analyses the figures. Spending rose 1.65 per cent from September, The Decoder reported.

Spending concentrated among a few

Among companies paying for AI, median spending in October was $16.32 per employee per month, 93 cents more than in September, according to Ramp’s page. For the top 1 per cent of spenders, the median was $10,199.20, about 9 per cent more than in September.

Kharazian said 1 per cent of companies drive 80 per cent of revenue for OpenAI and Anthropic. The Neuron added that the figure refers to revenue visible in Ramp’s data. The top group consists mostly of venture-backed technology, finance and technical services companies, The Decoder reported.

In October, 56.8 per cent of Ramp’s business customers paid for some AI service, 1 percentage point more than in September. The share paying Anthropic was 44.8 per cent and the share paying OpenAI 40.6 per cent; a company can pay both. Payments to OpenAI grew 2.1 per cent in the month and payments to Anthropic 0.5 per cent, The Decoder reported.

Estimates of sample bias point in opposite directions. The Decoder said the figures probably overstate adoption, because Ramp’s customers are larger, more tech-friendly and more growth-oriented than average. Ramp itself estimates that its figures are likely too low, because free tools and work done on personal accounts do not show up in payments.

Token prices fall

Language models process text as tokens, or parts of words, and companies that use models through an API usually pay for the tokens they use. The average price per million text tokens fell 21 per cent over the previous month, The Decoder reported. Ramp’s page shows the blended OpenAI and Anthropic price falling 14.8 per cent, from $0.62 in the week ending 4 October to $0.53 in the week ending 1 November. The sources do not explain the difference.

Over the same period, the number of tokens processed rose 29 per cent and the money spent on them 10.5 per cent. The token figures come only from users of Ramp’s Token Spend Management service. Ramp said they do not represent AI spending by all its customers, although the trend is similar. In The Decoder’s assessment, competition between OpenAI and Anthropic and Chinese open-weight models are pushing prices down.

Customer support gains share

Software development teams accounted for 72 per cent of AI spending classified by department in October, but their spending has fallen from its summer peak. The share of customer support and customer success rose from 0.8 per cent in June to 13.3 per cent. Kharazian suspects customer service chatbots are behind the rise. The Neuron noted, however, that payment data does not show which applications drive the growth.

The Decoder said such users prefer cheaper and faster models, so OpenAI’s and Anthropic’s margins on new customers may be lower. Even so, the most expensive models’ share of spending edged up in October to 69.5 per cent. Kharazian counts Opus, Fable, Sol and Astra among them.

What this is based on

1. US corporate AI spending hits a new high in October, but growth slows to 1.65 percent

   Publisher: the-decoder.com · published 11 November 2026 at 21:35 (UTC+02:00) · retrieved 12 November 2026 at 05:58 (UTC+02:00) · discovery channel: web

2. Ramp AI Index

   Publisher: ramp.com · source publication time unknown · retrieved 12 November 2026 at 05:58 (UTC+02:00) · discovery channel: web

3. AI Spending Boom: Most Businesses Spend Very Little

   Publisher: www.theneuron.ai · published 12 November 2026 at 04:26 (UTC+02:00) · retrieved 12 November 2026 at 05:58 (UTC+02:00) · discovery channel: web

4. Ramp: US business AI spending hit a record in October, but grew just 1.65%

   Publisher: devisdead.ai · published 11 November 2026 at 22:19 (UTC+02:00) · retrieved 12 November 2026 at 05:58 (UTC+02:00) · discovery channel: web

5. US corporate AI spending hits new high… · AGI Hunt

   Publisher: agihunt.info · published 11 November 2026 at 21:35 (UTC+02:00) · retrieved 12 November 2026 at 05:58 (UTC+02:00) · discovery channel: web

6. US Corporate AI Spending Hits New Peak with Slowing Growth in October 2026

   Publisher: kaneaijapan.com · published 11 November 2026 at 21:35 (UTC+02:00) · retrieved 12 November 2026 at 05:58 (UTC+02:00) · discovery channel: web

Cologne-based translation company DeepL unveiled its new speech model in a press release on Wednesday 11 November. Aure produces the translated speech in DeepL Voice, which translates speech into another language in real time.

The Decoder and Trending Topics reported that DeepL Voice has produced translations in the speaker’s own voice since September. That voice, however, was based on a sample heard at the start of the conversation. DeepL said the fixed sample conveyed who was speaking but not how that person spoke. The translation did not follow changes in emotion, emphasis or pace, nor sarcasm, where the tone of voice contradicts the words.

Voice, prosody and content kept separate

Aure analyses speech throughout the conversation. The model decides how to deliver a sentence based on both the translated words and the original voice, the press release said. DeepL does not permanently store meeting transcripts or translations, so it has little data on the same speaker in different languages. The model must therefore adapt to the speaker mid-sentence.

In training, DeepL separated the speaker’s voice identity, prosody – the melody and rhythm of speech – and spoken content, so that each can be carried into the new language independently. The company used reinforcement learning to refine the model’s output so that translations stay accurate and speech sounds natural, without a foreign accent. The model also separates the speaker’s voice from background noise and echo.

South Korea’s Yonhap news agency wrote that Aure converts the audio directly rather than first transcribing speech into text. DeepL’s press release, however, said the model also uses the translated words, so the claim does not fully match the company’s own description.

Quality claims come from DeepL

The Decoder, heise online, Trending Topics and Yonhap base their reports on information from DeepL, and none of them presents an independent measurement of Aure’s quality.

Trending Topics cites a blind test commissioned by DeepL and conducted by the industry publication Slator. In it, 96 per cent of the linguists surveyed preferred DeepL Voice to the native translation tools from Google, Microsoft and Zoom. The test concerned DeepL Voice, and the source does not say whether Aure was involved. DeepL’s product page uses a different metric: it says meeting translations contain 76 per cent fewer critical translation errors than the platforms’ own tools.

Trending Topics reported that Google Meet’s translation and Microsoft Teams’ Interpreter agent also aim to mimic the speaker’s voice.

Availability and privacy

Aure is enabled automatically in DeepL Voice for meetings and face-to-face conversations, and in DeepL’s speech translation API. DeepL Voice can be tried free of charge, and the meetings version works in Microsoft Teams, Zoom and Google Meet, the product page says.

The initially supported languages are English, German, Spanish, French, Italian, Japanese, Portuguese and Korean. For other languages, DeepL Voice still uses earlier models. DeepL promises more languages.

Aure neither creates nor stores a persistent voice profile, the product page says. Voice data is processed in memory and deleted when the session ends, and DeepL does not use meeting data to train its models.

What this is based on

1. DeepL wants real-time translation to finally capture sarcasm and mood shifts

   Publisher: the-decoder.com · published 11 November 2026 at 15:50 (UTC+02:00) · retrieved 12 November 2026 at 05:56 (UTC+02:00) · discovery channel: web

2. DeepL unveils Aure, its most lifelike Voice AI translation model yet

   Publisher: www.deepl.com · published 11 November 2026 · retrieved 12 November 2026 at 05:56 (UTC+02:00) · discovery channel: web

3. DeepL Aure: A.I. Voice Translation That Carries Emotion | Trending Topics

   Publisher: www.trendingtopics.eu · published 11 November 2026 at 11:35 (UTC+02:00) · retrieved 12 November 2026 at 05:56 (UTC+02:00) · discovery channel: web

4. KI-Stimmen mit Emotionen: DeepL stellt neues Modell Aure vor

   Publisher: www.heise.de · published 11 November 2026 at 15:52 (UTC+02:00) · retrieved 12 November 2026 at 05:56 (UTC+02:00) · discovery channel: web

5. Aure: Das natürlichste Sprachmodell von DeepL

   Publisher: www.deepl.com · source publication time unknown · retrieved 12 November 2026 at 05:56 (UTC+02:00) · discovery channel: web

6. DeepL wants real-time translation to finally capture sarcasm and mood shifts - The Decoder

   Publisher: www.translations-news.com · published 11 November 2026 at 15:53 (UTC+02:00) · retrieved 12 November 2026 at 05:56 (UTC+02:00) · discovery channel: web

7. 목소리에 실린 감정까지 실시간 통역…딥엘, 음성AI '오러' 공개 | 연합뉴스

   Publisher: www.yna.co.kr · published 11 November 2026 at 10:00 (UTC+02:00) · retrieved 12 November 2026 at 05:56 (UTC+02:00) · discovery channel: web

Researchers from Chinese universities and a research institute, the University of California, Berkeley, and Simple Agent Lab published a preprint on arXiv on 5 November. It has not been peer reviewed. They examined whether higher test scores from automatic evolution mean a better AI agent.

The methods evolve the harness: the program that steers the agent, gives the model its tools and runs its steps. They modify the harness’s prompts, skills or code based on test results. The model weights stay unchanged.

Every version lost tasks

The researchers evolved two harnesses with four methods (AHE, HyperAgents, A-Evolve and GEPA). In MiniSWE a method could also change the source code; in Codex, only the outer layers. The tasks came from τ³-Banking, which mimics customer banking, and the coding benchmark Terminal-Bench 2. The researchers compared each of the 16 evolved versions with the seed version over eight runs.

Reliability, the average success rate, improved in 11 versions. Coverage, the share of tasks solved at least once, improved in ten. Yet every version failed entirely on some tasks that the seed version had solved. Versions lost 2–10 tasks and gained 1–15 new ones. Some of the τ³-Banking changes reversed direction between the first and second halves of the runs, the authors said.

In the table, the version evolved with HyperAgents raised reliability by 5.2 percentage points even though it lost ten tasks. Figures from a single setting cannot show which method is better in general.

τ³-Banking, MiniSWE. Cost is the estimated running cost per task across eight runs. AGE is the authors’ own method.

Of the 11 versions that improved reliability, eight cost more per run than the seed version. A more expensive version can still work out cheaper if it needs fewer attempts. The MiniSWE version evolved with AHE solved 56.7 per cent of τ³-Banking tasks in four runs; the seed version solved 54.6 per cent in eight runs, at a higher cost.

In the run logs, failures involved, for example, a wrong date, the agent stopping work midway, or a solution that worked only in the agent’s own environment.

The highest score did not predict later progress

The highest-scoring version did not yield the largest mean improvement in further evolution in any of the four comparison pools. Targeted changes improved reliability more than general reminders in 15 of 16 cases. When changes were combined, however, their benefit could change.

When the researchers tried the versions on other benchmarks without further modification, coverage usually fell on general tasks but rose on coding tasks with the Codex harness.

Building on their findings, the authors created the EvoAudit workflow and used it to build AGE, a variant of A-Evolve. In the authors’ own tests, AGE improved reliability and coverage over A-Evolve in all four settings. The authors stress that neither test scores nor expert support for EvoAudit’s suggestions show that an agent is safe to use.

The study’s code is in the RSIHub repository on GitHub.

What this is based on

1. Beneath the Surface: What We Overlook When We Talk About Agent Evolution

   Wang, Xiaobo, Wang, Zimu, Ye, Siyu, Li, Jiaqi, Tang, Min, Bai, Jun, Shi, Yuling, Wang, Ruobing, Zhang, Jingjing, Wang, Chaofan, Wu, Kai, Qi, Mengnan, Cai, Kai, Gu, Xiaodong, Liu, Qi, Zheng, Zilong · Publisher: arxiv.org · published 5 November 2026 at 10:06 (UTC+02:00) · retrieved 12 November 2026 at 05:59 (UTC+02:00) · discovery channel: web

2. Paper page - Beneath the Surface: What We Overlook When We Talk About Agent Evolution

   Publisher: huggingface.co · source publication time unknown · retrieved 12 November 2026 at 05:59 (UTC+02:00) · discovery channel: web

3. Higher agent harness scores do not guarantee reliable, efficient and transferable improvement

   Publisher: sguys99.github.io · published 4 November 2026 at 21:00 (UTC+02:00) · retrieved 12 November 2026 at 05:59 (UTC+02:00) · discovery channel: web

4. Beneath the Surface: What We Overlook When We Talk About Agent Evolution

   Publisher: www.roboticscenter.ai · published 5 November 2026 · retrieved 12 November 2026 at 05:59 (UTC+02:00) · discovery channel: web