Anthropic released the Claude Fable 5.2 and Mythos 5.2 models on 9 November. At the same time, the company said the preserved thinking feature now applies to all API accounts. Its developer account, ClaudeDevs, explained on X what the change means for Fable 5.2 users.

Preserved thinking prevents anyone from editing Claude’s prior context. Distillers have edited that context to extract the model’s reasoning, Anthropic said. In distillation, a model is trained on the outputs of a more capable model so that it acquires that model’s capabilities.

Unauthorised distillation happens on an industrial scale, often through thousands of fake Claude accounts, Anthropic said. The company described Fable 5.2 as the best protected against distillation of all the models it has released so far. It said its countermeasures have reduced the data distillers collect by 90–99 per cent. The figures are the company’s own.

What the change requires of applications

ClaudeDevs said editing prior context can now result in a 400 error, meaning the API rejects the request. The change may affect harnesses that edit prior context. A harness is a program that controls an agent: it gives the model tools and runs its work steps.

Instead of editing history, Anthropic advises using compaction, the API’s context editing feature or mid-conversation system messages. Claude Code, Managed Agents and the Agent SDK handle this themselves. The company points developers updating their own integrations to its preserved thinking guidance.

Data retention and model choice

Fable 5.2 requires data retention by default so that Anthropic’s safety classifiers can operate, GitHub said. Zero data retention is available to certain eligible enterprise customers, GitHub added.

Anthropic said models this capable need monitoring that examines multiple requests together. The Enterprise Frontier Safeguards service automatically monitors prompts and responses stored in the customer’s own infrastructure. The service is in closed beta. Eligible customers can use the models with zero data retention even before they adopt the service. For other customers, Anthropic retains data for 30 days for monitoring.

ClaudeDevs also published a model selection guide. It explains when adjusting reasoning effort is enough and when to choose Fable 5.2, Opus 5.5, Sonnet or Haiku. Anthropic recommends starting most coding and knowledge work tasks with Opus 5.5 and using Fable for the most demanding, open-ended technical work.

Fable 5.2 costs $8 per million input tokens and $40 per million output tokens. Claude Code version 2.1.321 made it the default version of Fable. Saved Fable 5.1 selections switch to the new model immediately on Pro and Max plans, but on Team and Enterprise plans only once the organisation allows it. The command /model claude-fable-5-1 switches back to the old model.

What this is based on

1. Introducing Claude Fable 5.2 and Claude Mythos 5.2

   Publisher: www.anthropic.com · published 9 November 2026 · retrieved 10 November 2026 at 05:08 (UTC+02:00) · discovery channel: web

2. github.com

   Publisher: github.com · published 9 November 2026 at 21:22 (UTC+02:00) · retrieved 10 November 2026 at 05:08 (UTC+02:00) · discovery channel: GitHub

3. To help prevent distillation, preserved thinking now covers all API accounts on Fable 5.2. Editing prior context can 400. Use compaction, context editing, or mid-conversation system messages.…

   ClaudeDevs · Publisher: x.com · published 9 November 2026 at 21:04 (UTC+02:00) · retrieved 10 November 2026 at 04:00 (UTC+02:00) · discovery channel: X

4. Claude Fable 5.2 requires data retention by default to operate Anthropic's safety classifiers. Zero data retention access is available for certain eligible enterprise…

   github · Publisher: x.com · published 10 November 2026 at 00:53 (UTC+02:00) · retrieved 10 November 2026 at 04:02 (UTC+02:00) · discovery channel: X

5. Which Claude model should you build with? Our new guide breaks down when to use Fable 5.2, Opus 5.5, Sonnet or Haiku. It also covers when to adjust effort instead of switching models, and how to…

   ClaudeDevs · Publisher: x.com · published 9 November 2026 at 23:03 (UTC+02:00) · retrieved 10 November 2026 at 04:00 (UTC+02:00) · discovery channel: X

6. I really love using Fable 5.2. On big, open-ended work it feels like a major step above Opus 5.5. A few tips for getting the most out of it: - Ask what’s wrong with your idea before it builds…

   alexalbert__ · Publisher: x.com · published 9 November 2026 at 21:08 (UTC+02:00) · retrieved 10 November 2026 at 04:17 (UTC+02:00) · discovery channel: X

7. thenewstack.io

   Publisher: thenewstack.io · published 9 November 2026 at 21:08 (UTC+02:00) · retrieved 10 November 2026 at 05:08 (UTC+02:00) · discovery channel: web

Anthropic released Claude Fable 5.2 on 9 November. The same day, two external evaluators, Artificial Analysis and Epoch AI, posted their own results on X. The Artificial Analysis Intelligence Index combines nine benchmarks covering areas such as coding, scientific reasoning and knowledge work.

Fable 5.2 scored 65 on the index at its highest reasoning effort (max). Reasoning effort sets how much the model may think before answering, and Fable has five levels. The nearest other models were Gemini 4 Carbon (60) and Claude Opus 5.5 (59).

Artificial Analysis said the lead comes from scientific reasoning, knowledge and coding. The model led five of the nine benchmarks. In Artificial Analysis’s tests it scored 80 per cent on the HLE-Diamond reasoning benchmark (previous best 69 per cent), 73 per cent on Terminal-Bench-Science research tasks (65 per cent) and 49 per cent on the software development benchmark FrontierCode 2.0 (46 per cent). Anthropic’s own results on the last two were almost identical: 74.6 and 49.1 per cent.

The model did not lead every benchmark. On the AA-Briefcase knowledge-work benchmark, its 1,785 Elo points fall within the confidence interval of Opus 5.5’s score (1,800). On the Terminal-Bench 4.0 coding benchmark, it scored 60 per cent at max reasoning effort, the same as Opus 5.5, and trailed models such as Kimi K3.1 Pro (67 per cent).

The tests used the fallback model that Anthropic provides by default. Fable’s safety classifier blocks most cybersecurity tasks, among other things, Anthropic said. In those cases an older Opus model answers the query.

The top score is costly and slow

At max reasoning effort, one index task cost $26.54 on average. Artificial Analysis said that is about 2.6 times the cost of Opus 5.5 and about twice that of Fable 5.1, even though the per-token price fell 20 per cent from the predecessor. The model generated 275 million tokens during the index run, against a median of 72.8 million for models in the same price class. A task took 46.4 minutes on average (median 12.7).

At high and xhigh effort, Fable 5.2 scored 62, more than any other model at any level. At lower levels it can outperform Opus 5.5 while still costing less per task, Artificial Analysis said. The table shows index scores and cost and time per task.

Source: Artificial Analysis, Intelligence Index v5.0.1. Averages per index task; – = not published.

Gemini 4 Carbon was measured at high effort, Opus 5.5 and Fable 5.1 at max.

Epoch: new record, wide uncertainty

Epoch AI said Fable 5.2 raised the record on the Epoch Capabilities Index to 174. The previous record was GPT-6 Astra’s 169. The score exceeds what the trend for reasoning models before GPT-6 Astra would have predicted, Epoch said. In the chart Epoch published, however, the score’s 90 per cent uncertainty interval reaches into the trend’s prediction interval. Epoch tested the model before release using only its own benchmarks. The figure may change.

Fable 5.2 is available on Anthropic’s Claude Platform as claude-fable-5-2 and through AWS, Google Cloud and Microsoft Azure. It costs $8 per 1 million input tokens and $40 per 1 million output tokens. Anthropic recommends starting most coding and knowledge work with Opus 5.5.

What this is based on

1. Introducing Claude Fable 5.2 and Claude Mythos 5.2

   Publisher: www.anthropic.com · published 9 November 2026 · retrieved 10 November 2026 at 05:08 (UTC+02:00) · discovery channel: web

2. Claude Fable 5.2 Models - Intelligence, Performance & Price Comparison | Artificial Analysis

   Publisher: artificialanalysis.ai · source publication time unknown · retrieved 10 November 2026 at 05:11 (UTC+02:00) · discovery channel: web

3. Claude Fable 5.2 (max) - Intelligence, Performance & Price Analysis | Artificial Analysis

   Publisher: artificialanalysis.ai · source publication time unknown · retrieved 10 November 2026 at 05:08 (UTC+02:00) · discovery channel: web

4. Anthropic has released Claude Fable 5.2, taking the #1 spot on the Artificial Analysis Intelligence Index v5 at 65, led by gains in science, knowledge, and coding At max effort it scores 65, 5…

   ArtificialAnlys · Publisher: x.com · published 9 November 2026 at 21:45 (UTC+02:00) · retrieved 10 November 2026 at 04:14 (UTC+02:00) · discovery channel: X

5. Full breakdown of the individual evaluations in the Artificial Analysis Intelligence Index v5 for Claude Fable 5.2 across effort settings https://t.co/0VvlPCn6R1

   ArtificialAnlys · Publisher: x.com · published 9 November 2026 at 21:45 (UTC+02:00) · retrieved 10 November 2026 at 04:14 (UTC+02:00) · discovery channel: X

6. Compare Claude Fable 5.2 with other leading models at: https://t.co/8ho7ulef9p

   ArtificialAnlys · Publisher: x.com · published 9 November 2026 at 21:45 (UTC+02:00) · retrieved 10 November 2026 at 04:14 (UTC+02:00) · discovery channel: X

7. Claude Fable 5.2 sets a new record on the Epoch Capabilities Index (ECI) at 174. The previous high was GPT-6 Astra’s 169. Fable 5.2’s score is above what the pre-Astra trend for reasoning models…

   EpochAIResearch · Publisher: x.com · published 9 November 2026 at 23:22 (UTC+02:00) · retrieved 10 November 2026 at 04:13 (UTC+02:00) · discovery channel: X

Anthropic released version 2.1.321 of its coding agent Claude Code on GitHub on 9 November. The changelog does not say how serious the fixed flaws are or whether anyone has exploited them.

Permission and security fixes

When an MCP server redirected an OAuth sign-in request, Claude Code could send secrets to another server. These included the refresh token, which is used to obtain new access without signing in again.

The release also fixed three permission gaps:

• in one case, a read-only Bash command ran outside the sandbox without asking for permission
• in one case, the permission check for a Bash command evaluated a file write from the wrong directory and did not ask for permission
• deny rules and protected-file prompts did not apply to the file that cp or mv wrote into a directory.

cp and mv called by path, such as /bin/cp, now get the same path checks as plain cp and mv.

On the other hand, Claude Code now asks for permission less often. In auto and bypass modes, it enables mods that Claude writes without asking, even if the organisation has set allowUnsandboxedCommands: false or allowManagedDomainsOnly. The mods’ own tool calls trigger no permission prompt, even when the user’s own ask rules would cover them. Deny rules and organisation policies still apply.

Remote Control gains a --settings flag: the settings file it names applies to every session that claude remote-control starts. If the file changes, the command starts no new sessions until it is restarted. The Remote Control client can no longer move a session to another folder.

Fable 5.2 becomes the default

The default Fable model is now Claude Fable 5.2, which Anthropic announced the same day. Its context window is 1 million tokens. Input costs $8 and output $40 per million tokens, and cache reads $0.20 per million. Saved Fable 5.1 selections switch immediately on Pro and Max plans, but on Team and Enterprise plans only once the organisation has allowed the switch. The command /model claude-fable-5-1 restores the old model.

In its announcement, Anthropic said Fable 5.2 produces responses up to 30 per cent faster than Fable 5.1. Even so, the company recommends starting most coding tasks with Opus 5.5. Fable’s safeguards block most cybersecurity tasks, and Opus 4.8 answers blocked queries. Independent evaluator Artificial Analysis ranked Fable 5.2, run at its highest reasoning effort, first in its index among 74 models in the same price class. Within that price class, it was the most expensive per index task ($26.54) and nearly the slowest (46.4 minutes).

Fable 5.2’s fast mode generates tokens up to 2.5 times faster and costs $16 per million input tokens and $80 per million output tokens, Anthropic said. Version 2.1.322, released the next day, brings fast mode to Google Cloud’s Vertex in the global, us and eu regions, but the changelog does not say which models it covers. If Vertex does not allow it, Claude Code uses standard speed.

The Bedrock, Vertex and Foundry connections in Claude apps gateway no longer read the ANTHROPIC_*_BASE_URL environment variables; a different address now goes in the connection’s base_url setting. In policy_mode: layered, the gateway offers more managed MCP servers after the update. Other fixes cover Remote Control, hooks and cloud sessions, among other things.

What this is based on

1. v2.1.321

   anthropics · Publisher: github.com · published 9 November 2026 at 21:22 (UTC+02:00) · retrieved 10 November 2026 at 04:25 (UTC+02:00) · discovery channel: GitHub

2. github.com

   Publisher: github.com · published 10 November 2026 at 04:29 (UTC+02:00) · retrieved 10 November 2026 at 05:11 (UTC+02:00) · discovery channel: GitHub

3. Introducing Claude Fable 5.2 and Claude Mythos 5.2

   Publisher: www.anthropic.com · published 9 November 2026 · retrieved 10 November 2026 at 05:08 (UTC+02:00) · discovery channel: web

4. Claude blog (claude.com)

   Claude blog (claude.com) · Publisher: claude.com · source publication time unknown · retrieved 10 November 2026 at 04:08 (UTC+02:00) · discovery channel: web

5. Release: llm-anthropic 0.31

   Publisher: simonwillison.net · published 10 November 2026 at 01:58 (UTC+02:00) · retrieved 10 November 2026 at 05:09 (UTC+02:00) · discovery channel: web

6. Anthropic launches new Claude Fable and Mythos 5.2 models

   Publisher: www.theverge.com · published 9 November 2026 at 21:00 (UTC+02:00) · retrieved 10 November 2026 at 05:08 (UTC+02:00) · discovery channel: web

7. Claude Fable 5.2 (max) - Intelligence, Performance & Price Analysis | Artificial Analysis

   Publisher: artificialanalysis.ai · source publication time unknown · retrieved 10 November 2026 at 05:08 (UTC+02:00) · discovery channel: web

Meta Superintelligence Labs released Muse Ember on 9 November. The company said it built the model for tool calls, coding subtasks, request routing, data extraction and processing long documents.

Meta proposes a division of labour in which the larger Muse Wonder plans and Ember handles a large volume of repetitive steps. Ember processes text, images, audio and video. It has a context window of 1 million tokens.

Independent evaluation: same index, different strengths

Artificial Analysis (AA) builds its index from nine benchmarks, according to a chart it published. AA also calculates how much a single test task costs at a model’s list prices. The table compares Ember with Anthropic’s Claude Haiku 5.5 and Meta’s larger models.

Source: Artificial Analysis, Intelligence Index v5. Spark 1.3 figures from Meta’s announcement, which cites AA.

The same index score does not mean the same capabilities: Ember and Haiku do well in different component tests. AA said Ember’s token prices are the same as those of Haiku 5.5 and OpenAI’s GPT-6 Luna. TechCrunch disagrees, reporting that Ember is considerably cheaper than both on input token prices.

In AA’s coding evaluations, Ember beat Haiku on Terminal-Bench 4.0 (40.9 per cent vs 32.8 per cent) but lost on FrontierCode 2.0 (27.2 per cent vs 34.9 per cent). Meta’s table gives Haiku’s Terminal-Bench score as 39.2 per cent. Meta tested each model in its best published harness, Haiku in Claude Code. A harness is a program that controls an agent, giving the model tools and running its work steps.

On AA’s coding agent index, Ember scored 46 in Meta’s Muse Code coding agent, Haiku 37 in Claude Code and Luna 31 in Codex.

The sources give different figures for AutomationBench-AA, which measures business process automation. Meta’s announcement, which lists AA as its source, gives Ember 37.8 per cent under strict pass@1 scoring: a task counts as solved only on the basis of the first attempt. AA’s own announcement puts the score at 68.1 per cent, the same as Muse Wonder. No explanation is given for the difference.

AA identifies knowledge as Ember’s weakness. On AA-Omniscience, Ember answered 21.1 per cent of questions correctly, against 34.7 per cent for Haiku. AA said this was mainly because Ember declined to answer 64 per cent of the questions. As a result, its hallucination rate stayed at 17.4 per cent, half Haiku’s figure.

In Meta’s own evaluations, Ember led Haiku and Luna on benchmarks such as Agents’ Last Exam and OSWorld 2.1.

Price and availability

Ember is available as muse-ember-1.0 in the Meta Model API and in Muse Code. For prompts of up to 200,000 tokens, 1 million input tokens cost $0.10, cached input $0.004 and 1 million output tokens $0.50. Prices triple for longer prompts. Meta said the model is also available on Google Cloud and Oracle Cloud.

What this is based on

1. Introducing Our Smallest, Most Efficient Model: Muse Ember

   Publisher: research.meta.ai · published 9 November 2026 at 02:00 (UTC+02:00) · retrieved 10 November 2026 at 05:08 (UTC+02:00) · discovery channel: web

2. Meta

   Meta · Publisher: ai.meta.com · source publication time unknown · retrieved 10 November 2026 at 04:07 (UTC+02:00) · discovery channel: web

3. Meta releases Muse Ember, a new lightweight model for agent execution | TechCrunch

   Publisher: techcrunch.com · published 9 November 2026 at 20:00 (UTC+02:00) · retrieved 10 November 2026 at 05:11 (UTC+02:00) · discovery channel: web

4. Meta launches Muse Ember, a small AI model built for the repetitive work agents actually do

   Publisher: the-decoder.com · published 9 November 2026 at 21:41 (UTC+02:00) · retrieved 10 November 2026 at 05:12 (UTC+02:00) · discovery channel: web

5. Meta、最小・低コストのAIモデル「Muse Ember」発表 「Muse Spark 1.3」に迫る性能を約17分の1のコストで

   Publisher: www.itmedia.co.jp · published 10 November 2026 at 11:06 (UTC+02:00) · retrieved 10 November 2026 at 05:12 (UTC+02:00) · discovery channel: web

6. Meta's newest model Muse Ember scores 36 on the Artificial Analysis Intelligence Index, matching Claude Haiku 5.5 (max) at a third of the Cost per Task Muse Ember is Meta's newest efficient-tier…

   ArtificialAnlys · Publisher: x.com · published 9 November 2026 at 20:07 (UTC+02:00) · retrieved 10 November 2026 at 04:14 (UTC+02:00) · discovery channel: X

7. Full Artificial Analysis Intelligence Index results for Muse Ember (max) https://t.co/Kox8YHddVR

   ArtificialAnlys · Publisher: x.com · published 9 November 2026 at 20:07 (UTC+02:00) · retrieved 10 November 2026 at 04:14 (UTC+02:00) · discovery channel: X

On 10 November, the Ukrainian news agency UNN, citing Bloomberg, reported that Fireworks AI had disclosed its second large-scale cyberattack in a few weeks.

Fireworks AI is a US start-up whose service lets companies deploy AI models. Its investors include Nvidia and AMD, and it has more than 10,000 customers.

The latest incident took place on 24 October, Bloomberg’s sources said. An unidentified third party broke into the systems of the cloud provider that hosts Fireworks AI’s servers. The attackers gained access to tens of thousands of files, some of which contained customer data. UNN’s headline puts the number of files at thousands.

Fireworks AI has notified affected customers. The company says it is working with Notion, Cursor and other partners to limit the consequences of the attack. The attackers behind the latest breach have not yet been identified.

The company disclosed the previous incident in early October. In that case, the attackers used Fireworks AI’s internal credentials to access customer accounts. A subsequent investigation found that they had gained access to payment information and, in some cases, to confidential data, datasets and partly finished AI models.

Fireworks AI has brought in security specialists from Palo Alto Networks to help with the investigation and has notified law enforcement agencies.

The company was valued at $17.5 billion in July. Bloomberg reported that it is negotiating a funding round that could raise its valuation to $30–40 billion.

What this is based on

1. Nvidia-backed startup Fireworks AI hacked for the second time in several weeks, hackers gained access to thousands of files

   Publisher: unn.ua · published 10 November 2026 at 03:37 (UTC+02:00) · retrieved 10 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

NVIDIA released the Nemotron 3.5 Bolt 74B-A9B 24GB model on Hugging Face on 15 October. Sam Witteveen, who tests AI models on YouTube, walked through how the model was made in a video published on 9 November.

Bolt is an open-weight mixture-of-experts (MoE) model: for each token it uses only some of its subnetworks, which are called experts. Of its 74 billion parameters, 9 billion are active per token. NVIDIA designed the model for AI agents that run locally and sized it to fit in the memory of the RTX Spark device.

Quantisation stores a model’s weights with fewer bits. Witteveen said rounding the weights of a finished model usually works down to four bits, but quality drops sharply between two and three bits.

Pruning first, quantisation last

According to the model card, NVIDIA pruned Bolt from the 120-billion-parameter Nemotron 3.5 Super VL model. An architecture search chose the number and width of experts layer by layer so that the quantised model fits in under 24 gigabytes with a 128,000-token context window. The number of experts used per token fell from 22 to nine.

The pruned model recovered its capabilities by learning from the parent model. NVIDIA then fine-tuned it and trained it with reinforcement learning. Finally, the company taught the quantised model to imitate its unquantised version. Witteveen said this is how the model learns to work at a low bit count.

NVIDIA stored all experts at 2.0625 bits per weight. The Mamba layers use the 4-bit NVFP4 format and the embedding and output layers are 8-bit, so the model is not entirely 2-bit.

NVIDIA cites earlier research, such as the ParetoQ and Quest preprints, to support the approach. According to these preprints, layers twice the size can perform better at two bits than the originals at four. Witteveen remains sceptical, because outsiders find it hard to replicate this kind of training.

The benchmark figures are NVIDIA’s own

In most of the 13 benchmarks on NVIDIA’s model card, the 2-bit Bolt scores higher than three smaller MoE models measured unquantised in BF16 format. The table shows five of the benchmarks.

Bolt does not come out on top in every benchmark in the table: Qwen3.6 is narrowly ahead in PinchBench, and some of the comparison models score higher on both AA-Omniscience metrics. According to NVIDIA figures presented by Witteveen, quantisation lowered the PinchBench score from about 90 to 88 and the GPQA Diamond score from about 89 to 86. Witteveen said the unquantised version takes up 148 gigabytes.

Witteveen: smarter, not faster

Witteveen said NVIDIA’s speed table puts Bolt at around 84 tokens per second on the RTX Spark, fewer than Qwen 3.6 and Gemma 4. Witteveen described Bolt as a model that is smarter per gigabyte of memory but not faster.

In Witteveen’s own test, Bolt scored 10/10 on basic tool-call tests. It sometimes failed to fend off prompt injection, as did Super. In Witteveen’s view, Bolt suits Spark-class devices, but on a more powerful GPU a larger model may be the better choice.

The model can be downloaded from Hugging Face under the OpenMDW-1.1 licence, which permits commercial use. A GGUF version is also available. The model card’s instructions use the development version of vLLM or newer. The model card lists the supported GPUs differently in different places: either A100, H100 and Blackwell GPUs, or only Blackwell’s SM120 and SM121 chips.

What this is based on

1. nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB

   Publisher: huggingface.co · source publication time unknown · retrieved 10 November 2026 at 05:49 (UTC+02:00) · discovery channel: web

2. www.youtube.com

   Publisher: www.youtube.com · published 9 November 2026 at 16:00 (UTC+02:00) · retrieved 10 November 2026 at 05:49 (UTC+02:00) · discovery channel: YouTube

ARC Prize published Claude Fable 5.2’s ARC-AGI results on 9 November. The same day, Anthropic announced on X that Fable 5.2 and Claude Mythos 5.2 are available. ARC Prize tested Fable 5.2 with two harnesses at five reasoning effort levels, ten configurations in total.

A harness is the program that controls an agent: it gives the model its tools and runs its steps. ARC-AGI-3 tasks are games, and a single game can last up to the 12-hour time limit. In ARC Prize’s Standard harness, the model can keep notes of its own choosing for the whole game.

The Provider Adapter harness instead preserves the model’s opaque reasoning state between requests. In long conversations it uses compaction. This lets the model reuse its prior work, ARC Prize said.

The gap appears at every reasoning effort level

The table shows ARC-AGI-3 results with both harnesses. The best scores, 81.2 per cent and 97.7 per cent, came at different reasoning effort levels, so they cannot be compared directly. At the same XHigh level, the gap is 16.1 percentage points.

Source: ARC Prize, verified results on the semi-private test set, 9 November 2026.

ARC Prize does not say which Provider Adapter feature explains the gap: preserving reasoning state or compaction. This suggests that comparing ARC-AGI-3 scores requires knowing which harness was used.

At Max, the Standard harness reached only 65.7 per cent. In 28 of 55 games the run lasted the full 12-hour time limit, ARC Prize said. In those games, the time limit constrained the test’s progress.

The gap also shows in individual public demo games. In TU93 and BP35, the Standard harness scored 2.2 per cent at XHigh, while Provider Adapter scored 100 per cent. Fable 5.2’s best results at XHigh were 98.5 per cent on ARC-AGI-1 and 93.8 per cent on ARC-AGI-2.

Artificial Analysis: top of the index, but not in every benchmark

Benchmarking company Artificial Analysis said on X that Fable 5.2 at Max took first place in its Intelligence Index v5 with 65 points. The closest other models are Gemini 4 Carbon (High, 60) and Claude Opus 5.5 (Max, 59). On the Terminal-Bench 4.0 coding benchmark, Fable 5.2 scored 60 per cent, level with Opus 5.5. Kimi K3.1 Pro scored 67 per cent.

At Max, an index task cost an average of $26.5, about 2.6 times the cost for Opus 5.5 ($10.0), Artificial Analysis said. At High and XHigh, Fable 5.2 scored 62 points, more than any other model at any level.

Fable 5.2 costs $8 per 1 million input tokens and $40 per 1 million output tokens. Artificial Analysis said this is 20 per cent less than Fable 5.1’s price. The context window is 1 million tokens, and the model accepts text and images.

What this is based on

1. Claude Fable 5.2 - ARC-AGI Results

   Publisher: arcprize.org · published 9 November 2026 · retrieved 10 November 2026 at 05:20 (UTC+02:00) · discovery channel: web

2. Claude Fable 5.2 is now #1 on AA-3D, our benchmark for how well AI models build interactive 3D experiences Claude Fable 5.2 (Max) scores 1271 Elo, 67 points ahead of Claude Sonnet 5.5 (1204) and…

   ArtificialAnlys · Publisher: x.com · published 10 November 2026 at 02:41 (UTC+02:00) · retrieved 10 November 2026 at 04:14 (UTC+02:00) · discovery channel: X

3. Anthropic has released Claude Fable 5.2, taking the #1 spot on the Artificial Analysis Intelligence Index v5 at 65, led by gains in science, knowledge, and coding At max effort it scores 65, 5…

   ArtificialAnlys · Publisher: x.com · published 9 November 2026 at 21:45 (UTC+02:00) · retrieved 10 November 2026 at 04:14 (UTC+02:00) · discovery channel: X

4. Claude Fable 5.2 and Claude Mythos 5.2 are available now:

   AnthropicAI · Publisher: x.com · published 9 November 2026 at 21:00 (UTC+02:00) · retrieved 10 November 2026 at 04:23 (UTC+02:00) · discovery channel: X

Anthropic said in its announcement of Claude Fable 5.2 on Monday 9 November that it would give users of certain paid plans extra Fable usage during launch week. Plans and seats that include Fable models get the resets, the announcement said.

Claude plans cap usage with a weekly usage limit. Neither Anthropic’s announcement nor its X posts explain exactly how a reset works.

ClaudeDevs, the X account of Anthropic’s Claude Code team, posted the tiers for Ultra and Team Premium subscribers. Those who have subscribed for less than six months get one reset, those who have subscribed for 6–12 months get two and those who have subscribed for more than a year get three. Subscribers can use the resets at any time until 8 March, the post said.

Pro and Team Standard subscribers get one reset and a week to try Fable 5.2, ClaudeDevs said. During the trial, Fable usage can take up at most half of the weekly usage limit. Anthropic’s announcement gives the same limit.

Differing information on Max

Boris Cherny of Anthropic wrote on X that the company wanted to give more to its longest-standing subscribers. An image in Cherny’s post shows the same three-tier split as the ClaudeDevs post.

In another post, Cherny wrote that Pro users get one reset and Max, Ultra and Team Premium subscribers up to three. Neither the ClaudeDevs post nor Anthropic’s announcement mentions Max by name, so Max subscribers should check the benefit in their own account.

What to use Fable for

Anthropic said Fable 5.2 is best suited to the most ambitious, open-ended technical tasks. For most coding and knowledge-work tasks, the company recommends starting with Claude Opus 5.5 and switching models as needed.

Fable’s safeguards block most cybersecurity tasks and most biology research tasks. Opus 4.8 answers blocked cybersecurity requests and Opus 5 answers blocked biology requests, the announcement said.

In Claude Code version 2.1.321, the default Fable model is Fable 5.2. Saved Fable 5.1 selections move to the new model immediately on Pro and Max plans, and on Team and Enterprise plans only once the organisation allows it. Users can switch back to the old model with the command /model claude-fable-5-1.

What this is based on

1. Introducing Claude Fable 5.2 and Claude Mythos 5.2

   Publisher: www.anthropic.com · published 9 November 2026 · retrieved 10 November 2026 at 05:08 (UTC+02:00) · discovery channel: web

2. github.com

   Publisher: github.com · published 9 November 2026 at 21:22 (UTC+02:00) · retrieved 10 November 2026 at 05:08 (UTC+02:00) · discovery channel: GitHub

3. We're giving Ultra and Team Premium users Fable resets based on how long you've subscribed: <6 mo: 1 6–12 mo: 2 >12 mo: 3 Use them anytime by March 8. Pro and Team Standard get a week to…

   ClaudeDevs · Publisher: x.com · published 9 November 2026 at 21:04 (UTC+02:00) · retrieved 10 November 2026 at 04:00 (UTC+02:00) · discovery channel: X

4. Pro users get one reset, and Max, Ultra, and Team Premium get up to three, depending on how long you've subscribed: https://t.co/JkknwZZGyh

   bcherny · Publisher: x.com · published 9 November 2026 at 22:32 (UTC+02:00) · retrieved 10 November 2026 at 04:07 (UTC+02:00) · discovery channel: X

5. x.com

   ClaudeDevs · Publisher: x.com · published 9 November 2026 · retrieved 10 November 2026 at 05:22 (UTC+02:00) · discovery channel: X

6. We wanted to do more for the people who’ve been with us the longest. So with Fable 5.2, the longer you’ve been subscribed, the more Fable resets you get. https://t.co/g2Y5cwS0L2

   bcherny · Publisher: x.com · published 9 November 2026 at 21:21 (UTC+02:00) · retrieved 10 November 2026 at 04:07 (UTC+02:00) · discovery channel: X

7. Anthropic launches new Claude Fable and Mythos 5.2 models

   Publisher: www.theverge.com · published 9 November 2026 at 21:00 (UTC+02:00) · retrieved 10 November 2026 at 05:08 (UTC+02:00) · discovery channel: web

OpenAI announced Control Room on its blog and X account on 9 November. The company said the system covers every research inference workload, meaning every run in which a model produces output. Examples include agentic sampling for model training and evaluation, and data processing. Both posts are the company’s own and mention no external evaluation.

OpenAI has deployed Control Room in its research infrastructure, not in the model service it offers customers. The company said that service is more centralised and already has an established way to withdraw models quickly.

Permission for every model call

When a research job starts, Control Room gives it an identifier. Each time the job requests more tokens from a model, it passes the identifier to the inference engine, the software that runs the model on GPUs. The engine asks Control Room for permission before it responds. If the job has been stopped, the engine denies the request, and the job can no longer get output from any model on any engine.

OpenAI placed the check in inference because the code in researchers’ jobs varies and changes often, whereas inference has very few code paths and they are uniform. The company also uses agents and programmatic checks to search continuously for engines where the check is missing or switched off.

For example, a monitoring system can decide to stop a job when a monitor – a second model that watches the model being trained – raises an alert. OpenAI said a monitor watches every reinforcement learning run of its frontier models. The monitor also detected the DNS-related incident the company recently reported.

OpenAI also requires every research job that runs agents to have an approved safety case, a written explanation of why the job is safe enough.

35-second latency and failures

The company said a job stops within 35 seconds of Control Room receiving a stop instruction. The latency comes from caching permission checks, which the high volume of requests makes necessary. OpenAI said it aims to cut the latency to zero. The introduction to the blog post and the X post count the same 35 seconds from the decision to intervene, so they do not make clear how long it takes to pass the decision on.

The company said it built the system to keep working even when other systems fail. If Control Room crashes or someone disables it, for example through sabotage, the engines deny all inference requests until the service recovers. OpenAI said this safeguard covers all frontier model research jobs and will extend to other research jobs in the coming days.

What this is based on

1. Control Room: A central system to track and stop model workloads across our research infrastructure

   Publisher: openai.com · published 9 November 2026 at 15:00 (UTC+02:00) · retrieved 10 November 2026 at 05:25 (UTC+02:00) · discovery channel: web

2. We built Control Room, a central system to track model workloads across our research infrastructure and stop them quickly when safety or security concerns arise. It connects every research…

   OpenAI · Publisher: x.com · published 10 November 2026 at 00:01 (UTC+02:00) · retrieved 10 November 2026 at 04:24 (UTC+02:00) · discovery channel: X

Nous Research released Hermes Agent 0.23.1, a patch release, on GitHub on 9 November. Hermes Agent is Nous’s open-source AI agent, a program that carries out tasks using a language model. The release notes say the version rolls up about 618 pull requests merged since 0.23.0 into a stable release. It ships to the CLI, the installers, the Microsoft Store app, Docker and Hermes Cloud.

Hermes fetches its plugin catalogue from the web, and the hermes plugins install command looks up plugins in it. When Nous updated the catalogue’s site on 9 November, the catalogue file grew beyond 2 MiB. Installs do not accept a larger file, so every install silently fell back to the catalogue that shipped with its version.

As a result, the command could not find plugins added since the user’s last update, the release notes say. It also installed the pinned plugin version that shipped with the user’s Hermes version instead of the current one, and ignored newer removal entries in the catalogue.

Version 0.23.1 raises the cap to 16 MiB. Versions 0.23.0 and older keep using the old catalogue until they are updated. The CLI updates with hermes update, and the Microsoft Store app through the Store. On Docker and Hermes Cloud, the release is in the Docker image nousresearch/hermes-agent:stable. The Desktop app installed from the installer (.exe or .dmg) does not update in this release, only with the next full release.

The release notes deliberately leave out other changes, such as cancelling with the Esc key in the text interface and new plugin APIs. Full release notes are promised for version 0.24.0.

Cloud agents open in the browser

Nous Research said on X the same day that Hermes Cloud agents now open in a browser tab with the same interface as the Hermes Desktop app. Users create an agent in Nous Portal by choosing a name and a model. Nous hosts the agent in a separate container around the clock, so it can run scheduled tasks and reply in messaging channels even when the user’s computer is off.

Chats, group chats, plugins and settings work in the browser as they do on a local machine, the company said. Users can also use the agent on a phone or connect Hermes Desktop to it.

The Plus, Super and Ultra plans ($20, $100 and $200 a month) include one cloud agent. On Plus the agent has up to 4 GB of memory and 25 GB of disk space; on Ultra, 8 GB and 100 GB. Without a plan, setting up an agent requires a balance of at least $2. The service’s page says the instance is charged daily to the Nous balance based on actual usage, and inference (running the model) and tools are billed separately.

Fable 5.2 also available with a Claude subscription

Teknium of Nous Research said on X on 9 November that the new Fable 5.2 model is available in Hermes Agent through Nous Portal and OpenRouter, and with a Claude subscription.

What this is based on

1. Hermes Agent v0.23.1

   NousResearch · Publisher: github.com · published 10 November 2026 at 03:42 (UTC+02:00) · retrieved 10 November 2026 at 04:28 (UTC+02:00) · discovery channel: GitHub

2. Hermes Cloud — Always-On AI Agent Hosting

   Publisher: portal.nousresearch.com · source publication time unknown · retrieved 10 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

3. Hermes Cloud agents now open in a browser tab with the exact same interface as Hermes Desktop. https://t.co/4Vw3XJgFwn To launch a personal agent, pick a name and a model in Nous Portal. We host…

   NousResearch · Publisher: x.com · published 9 November 2026 at 19:55 (UTC+02:00) · retrieved 10 November 2026 at 04:12 (UTC+02:00) · discovery channel: X

4. Try Fable 5.2 in Hermes Agent https://t.co/1b9u2lVCou

   NousResearch · Publisher: x.com · published 10 November 2026 at 00:57 (UTC+02:00) · retrieved 10 November 2026 at 04:12 (UTC+02:00) · discovery channel: X

5. New Fable 5.2 available in Hermes Agent through Nous Portal, OpenRouter, and Claude subs available 😉

   Teknium · Publisher: x.com · published 10 November 2026 at 00:23 (UTC+02:00) · retrieved 10 November 2026 at 04:15 (UTC+02:00) · discovery channel: X

6. With Hermes it's all yours to conquer https://t.co/whqZD3Y2Qk

   NousResearch · Publisher: x.com · published 10 November 2026 at 02:46 (UTC+02:00) · retrieved 10 November 2026 at 04:12 (UTC+02:00) · discovery channel: X

Google has released as open source a plugin for coding agents that it developed internally, the Korean outlet AI Times reported on 9 November. Thousands of Google engineers use it in product development, AI Times said.

The plugin is called Elephant-Goldfish Model (EGM) and is licensed under Apache 2.0. A description and installation instructions are in Google’s GitHub repository, which states that EGM is not an officially supported Google product.

The elephant plans, the goldfish review

The Elephant is a session that knows the project, its requirements and earlier decisions. It works through the problem with the user and writes a plan: a software specification document. The Elephant does not write code, AI Times reported.

The Goldfish are subagents that receive the plan and the source material it references, but not the earlier conversation. The repository says this helps expose blind spots and unstated assumptions. AI Times gave the rationale: when the same AI both plans and reviews, it can accept its own flawed initial assumptions.

The review has five stages. The reviewers assess the plan’s design decisions, clarity, fit with existing code and implementation plan, among other things, the repository says. The user decides which findings to fix. After approving the plan, the user asks the coding agent to implement it.

The plugin also has dedicated functions for brainstorming, working through review findings and documenting existing code. Users can compare documents stored in the repository with the current source code using the Drift report function.

No published measurements of the benefit

No benchmarks have been published showing that EGM improves accuracy, AI Times reported. The repository says the review does not replace tests or human judgement. The ready-made workflow is designed for software specifications. To review other documents, users choose the reviewers and standards themselves.

Runs inside an existing agent

EGM is not a model but a plugin for a harness: the program that steers an agent, gives the model its tools and runs its steps. EGM uses the user’s existing large language model and counts against the user’s usage limit, AI Times reported.

The Goldfish review requires a harness that can create new subagents without prior context. Without one, EGM stops the review. The repository includes installation instructions for Claude Code, Codex CLI, Antigravity and the ChatGPT and Codex desktop apps, among others.

What this is based on

1. github.com

   Publisher: github.com · source publication time unknown · retrieved 9 November 2026 at 09:26 (UTC+02:00) · discovery channel: GitHub

2. 구글, 코딩 전 설계 오류 잡는 '코끼리-금붕어' 플러그인 공개 - AI타임스

   Publisher: www.aitimes.com · published 9 November 2026 at 11:03 (UTC+02:00) · retrieved 10 November 2026 at 05:34 (UTC+02:00) · discovery channel: web

Chetas Joshi, an engineer on Robinhood’s data platform team, explained how the team reviews changes to its data pipelines in a talk published on the AI Engineer channel on 10 November 2026. Coding agents open pull requests – proposed changes awaiting approval – faster than the team’s few experts can read them. The changes affect schemas, configuration and data access permissions, among other things.

Scripts check the facts, an agent assesses the risks

Scripts check each change against data from production systems. They compare the schema with the schema registry and fetch the table owner and personal data tags from the data catalogue. They also look up the column’s readers in the lineage tool, which records which jobs read which data. A performance test runs the change in a test environment at production data volume. If a check fails, the pull request is halted.

At the same time, an AI agent reads the change and can fetch files from the repository, but it cannot access production systems. Rules tell the agent what the scripts already cover. They direct it to look for problematic query patterns, access permissions on sensitive data and readers missing from the lineage tool. The agent comments on individual lines, giving its reasons, scores the risk and either approves the change or passes it to a person.

In Joshi’s example, a column’s integer type was widened to allow larger values (integer → bigint). The scripts’ checks passed, but the agent found an unregistered job in the repository. The job wrote the value to a column in another table that still accepted only smaller integers. The write would have failed on the first value that was too large. A person confirmed that the job was in use, and the other table was fixed before the merge.

The agent cannot see manually run queries or dashboards, so it must not claim that there are no readers. It may only state that it found none in the places it searched. A person always handles access permissions on sensitive data, migrations that change the meaning of data and uncertain assessments.

The figures are Robinhood’s own

As of 31 August 2026, about 60 per cent of data pipeline pull requests were merged without human review, Joshi said. Manual review effort fell by about 75 per cent, saving about 132 working hours a month. Incidents linked to data pipeline changes fell by about 30 per cent. Joshi did not give a comparison period in the talk.

Joshi put the rate of incorrect approvals at about 2 per cent; these stemmed mainly from gaps in the rules. The team has no clear figure for unnecessary handovers to people. The on-call engineer updates the rules by hand once a week based on people’s feedback. The system does not learn automatically.

The scripts and the agent cannot see each other’s results

Joshi said the team made a mistake by building the scripts and the agent as separate systems that run in parallel. Neither sees the other’s output, and the agent cannot call the scripts. The team is exploring a design in which a classifier decides how an entire pull request is handled and the agent gets the script checks as tools. This has not yet been built.

What this is based on

1. www.youtube.com

   Publisher: www.youtube.com · published 10 November 2026 at 03:00 (UTC+02:00) · retrieved 10 November 2026 at 05:47 (UTC+02:00) · discovery channel: YouTube

Baidu’s ERNIE team described its automated research system, ERNIE Evolve, on its research blog on 9 November. The company calls the system its first step towards recursive self-improvement (RSI), in which AI improves the software used to develop AI.

Agents experiment, a separate evaluator decides

The system takes a repository, an optimisation target and evaluation criteria as input. A Director agent chooses the research direction and starting version based on earlier experiments and a knowledge base built from research papers and open-source code. In parallel research lines, the Selector picks the code version and hypothesis to test, and the Executor modifies the code in an isolated environment.

The Evaluator and the Reviewer decide whether a change is accepted. The Evaluator is an immutable measurement program outside the repository and is not visible to the agents. An improvement is accepted only if it exceeds measurement noise. The Reviewer rejects changes that exploit loopholes in the evaluation (reward hacking).

Small margins over the leaders on public tasks

Baidu tested the system on three public tasks: improving a training algorithm with a five-minute training budget on a single GPU (NanoChat AutoResearch), speeding up training (NanoGPT Speedrun) and optimising GPU kernels (SOL-ExecBench). The table shows the results Baidu reported.

Source: Baidu. The NanoChat comparison is the result reported by rekursiv.ai; the others come from official leaderboards.

MarkTechPost notes that the margins are small and that Baidu reported the results itself. On SOL-ExecBench, ERNIE Evolve fell short of the top result.

Sources give different figures for competitors’ results. In Baidu’s measurement (Evaluation Stack v1.1), Recursive scored 0.676 on SOL-ExecBench and Tencent’s Hyra-1.0 scored 0.736, but MarkTechPost said the systems’ self-reported results are 0.754 and 0.771. The publication also said their Speedrun times of 77.5 and 76.4 seconds were compared with an older record of about 79.7 seconds, so they are not directly comparable.

ERNIE-5.5-Flash inference got faster

Baidu also used the system to optimise inference, or running the model, for ERNIE-5.5-Flash, which is still in development. Engineers had already optimised most of the baselines, and the system’s changes had to pass the same load tests and correctness checks as theirs. According to Baidu:

• On the Kunlunxin P800, the model was not yet making full use of existing optimised execution paths. The first round enabled them and raised generation speed to 282.9 per cent of the baseline; the fourth round raised it to 623.6 per cent.
• Reinforcement learning training requires deterministic inference: outputs had to stay bit-for-bit identical regardless of which other requests were processed in the same batch. Its throughput rose by 33.7 per cent in about three days, with determinism preserved.
• Single-request decoding speed on GPUs rose by 14.9 per cent with accuracy unchanged. Some useful changes sped things up by only 0.3–0.5 per cent on their own, which is below measurement noise.

MarkTechPost reported that Baidu has not released the system itself, only the optimised code and measurement logs for the public benchmark tasks on GitHub (ernie-research/ERNIE-Evolve-Results). Baidu next plans to use the method to optimise training systems. The company said RSI must stay under human control and remain interruptible.

What this is based on

1. ERNIE Evolve: Toward Automated Research for Real-World AI R&D · ERNIE

   Publisher: ernie.baidu.com · published 9 November 2026 · retrieved 10 November 2026 at 05:42 (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 10 November 2026 at 04:01 (UTC+02:00) · discovery channel: web

3. Baidu’s ERNIE Evolve Automated Research System Beats NanoChat and NanoGPT Records

   Publisher: www.marktechpost.com · published 9 November 2026 at 18:31 (UTC+02:00) · retrieved 10 November 2026 at 05:42 (UTC+02:00) · discovery channel: web

4. 百度文心首次披露RSI进展：ERNIE Evolve已用于最新模型研发，文心5.5 Flash 或将近期发布 | 雷峰网

   Publisher: www.leiphone.com · published 9 November 2026 at 11:18 (UTC+02:00) · retrieved 10 November 2026 at 05:42 (UTC+02:00) · discovery channel: web

5. Baidu’s ERNIE Evolve Automates AI Research and Inference Optimization

   Publisher: www.unite.ai · published 9 November 2026 at 21:14 (UTC+02:00) · retrieved 10 November 2026 at 05:42 (UTC+02:00) · discovery channel: web

Alibaba released version 2.0 of Qianwen Office (千问办公), its AI agent for office work, on 9 November. The Chinese technology outlets QbitAI and InfoQ published the company’s press release unchanged. The release describes it as the biggest update since the product launched about three months ago, and all the new features are already available in the Qianwen Office app.

The agent gathers information from several groups

With the Collaboration feature (协作), a team can set up a shared workspace for a project. Users can link DingTalk group chats, documents, spreadsheets and AI-written meeting notes to it. The agent, which the company calls a digital employee, uses these sources as background information. When a team member gives it a task, it drafts a status update, summarises what has been agreed, and creates and tracks tasks. The company said the feature is the first in the industry in which several people and an agent work together.

Pandaily reported that, according to the product documentation, a group administrator must approve the link when an ordinary member links a chat.

In an example from the release, five employees on the coupon team at the Intime department store (银泰百货) previously collected information from the group chats of more than 60 stores and entered it into a spreadsheet by hand. Now the groups and the spreadsheet are linked to the same workspace, and the agent gathers the information and updates the spreadsheet. The company said the workload fell significantly. Pandaily, however, noted that the example gives no measured time saving. Leiphone reported that Intime used the feature in an internal test. More than 200 people are involved in the work, and staff check about 50 details for each group-buying coupon.

Leiphone tried the feature itself: a reporter and an editor simulated brainstorming story ideas with the agent. The agent read interview material saved in the workspace and took both people’s comments into account. Leiphone nonetheless considers this a light use case compared with Intime’s multi-step work.

An event can trigger work

Users can choose an event and write a rule; when the rule is met, the agent starts work. The feature is called Automatic Tasks (自动任务). At first, the triggers are messages from a specified person, @-mentions of the user in a specified group, and approval tasks. Pandaily said the release notes also mention to-do list tasks, and that the feature extends an earlier scheduled-tasks capability. In the release’s example, a regional manager sets a rule: when someone @-mentions the manager in a group and says a product is out of stock, the agent reads the inventory spreadsheet and fills in a transfer request.

With the new co-editing feature, the agent edits the original file based on the user’s comments. Users can add several comments to a document, spreadsheet or slides. A new file space holds both files the agent produces and files the user uploads. Tasks run in the cloud and continue even when the computer is switched off. After granting permission, users can control their computer from their phone. Pandaily said enterprise administrators can switch off remote control.

Prices and user numbers have not been disclosed, and no independent test results on the agent’s accuracy in shared workspaces have been published, Pandaily reported.

What this is based on

1. 千问办公2.0发布，首个支持多人与Agent协作的Team Agent来了

   Publisher: www.qbitai.com · published 9 November 2026 at 10:34 (UTC+02:00) · retrieved 10 November 2026 at 05:42 (UTC+02:00) · discovery channel: web

2. 千问办公2.0，补上了办公Agent最关键的「短板」 | 雷峰网

   Publisher: www.leiphone.com · published 9 November 2026 at 16:16 (UTC+02:00) · retrieved 10 November 2026 at 05:42 (UTC+02:00) · discovery channel: web

3. Alibaba's Qianwen Office 2.0 Brings AI Agents Into Team Workspaces and Event-Triggered Tasks

   Publisher: pandaily.com · published 10 November 2026 at 05:18 (UTC+02:00) · retrieved 10 November 2026 at 05:42 (UTC+02:00) · discovery channel: web

4. 千问办公2.0发布，首个支持多人与Agent协作的Team Agent来了

   Publisher: www.infoq.cn · published 9 November 2026 at 10:26 (UTC+02:00) · retrieved 10 November 2026 at 05:42 (UTC+02:00) · discovery channel: web

ByteDance’s Seed research team introduced Seedream 5.5 in a blog post on 9 November. The new model generates and edits images for professional image production. The Decoder, which covers AI news, reported on the announcement the same day.

Repeated editing and transparent backgrounds

When someone edits an image in rounds, each building on the previous result, small errors can accumulate: details disappear and the image degrades. ByteDance said Seedream 5.5 was trained specifically to preserve unedited regions with pixel accuracy and without colour shifts. This supports continued work on the same image across 100 rounds of editing, the company said.

The company illustrated this with a demo in which it edited an image of the same person 100 times in succession. Even then, the model accurately edited small objects such as rings and glasses frames, the company said.

Second, the model stores each pixel’s transparency, known as the alpha channel. It produces RGBA images directly, with fully or partially transparent backgrounds. Users can place a generated object or character into a layout or composite without separate background removal or edge clean-up. Objects and characters can be generated from a text description or a reference image.

The company said the model also renders light, materials and skin more naturally, and draws small, dense text and multi-stroke Chinese characters more accurately than before.

The company’s own assessments

ByteDance tested the model in six use cases: e-commerce, marketing, film and television, games, office productivity and fashion design. Experienced image-generation evaluators scored the quality and usability of the results on the MagicArena platform.

Compared with the previous version, Seedream 5.0 Pro, both the evaluators’ quality scores (MOS) and the share of usable images rose in all six use cases. In film and television and in office productivity, the share of usable images more than doubled.

Experts assessed text structure and realism in image pairs using the GSB (Good/Same/Bad) method, in which the evaluator says which image is better or whether both are equally good. The new model beat the previous version in all four comparisons, across both text-to-image and image-to-image tasks. For realism in image-to-image tasks, for example, the GSB score was +55.62 per cent.

ByteDance said the model outperforms GPT Image 2.5 Sunburst and Nano Banana 2.1 in text structure in both task types. In realism, the company reports a clear lead in text-to-image tasks. The Decoder presents the comparison with competitors as ByteDance’s claim.

In a test of Chinese and English text, the severe text corruption rate was 0.4 per cent and the minor text corruption rate 1.7 per cent. The company said both rates are clearly lower than the previous version’s and among the best in a comparison of frontier models.

The Decoder reported that the model is gradually rolling out in ByteDance’s Jimeng and Doubao services, Doubao Work and the Volcengine Ark platform. ByteDance said a Pro version is coming and that it plans to expand the model’s use in agent workflows.

What this is based on

1. Seed News - ByteDance Seed Team

   Publisher: seed.bytedance.com · published 9 November 2026 · retrieved 10 November 2026 at 05:43 (UTC+02:00) · discovery channel: web

2. Bytedance's Seedream 5.5 promises pixel-stable image editing across 100 rounds without color drift

   Publisher: the-decoder.com · published 9 November 2026 at 20:50 (UTC+02:00) · retrieved 10 November 2026 at 05:43 (UTC+02:00) · discovery channel: web

On 9 November, as it launched its new model Claude Fable 5.2, Anthropic described an experiment in which the model designed electronic devices from prompt to prototype. Its partner was Diode Computers, which designs and manufactures circuit boards and has a commercial relationship with Anthropic.

The devices were a speakerphone that connects to a computer over USB-C, a home voice-assistant hub and a battery-powered virtual pet. For the speakerphone, the model received only a short description: microphones, a speaker and Bluetooth with an antenna on the circuit board. The model wrote the specifications, drew the schematics, placed the components on a four-layer board and routed the copper traces between them. It also designed a 3D-printable enclosure, wrote the firmware and compiled the bill of materials.

Diode manufactured the boards directly from the model’s files without any revision rounds, and its engineers assembled and tested the devices following the model’s instructions. Anthropic said all three worked without changes to the circuits and are still in working order, the speakerphone even after weeks of use. Diode steered the model mainly with short directives such as ‘clean up the schematic’ and ‘continue’.

The result does not show whether the devices would be viable as products. They are prototypes that have had no electrical validation, antenna tuning, thermal testing or certification, and they were not designed for mass production. The company does not know how they will hold up under continuous use. The model had no access to measuring instruments or manufacturing machines, so people did the physical work.

How it differs from earlier models

Anthropic said the biggest change was in geometric reasoning: placing components and routing traces. Earlier models placed supporting components too far from the circuits and rarely corrected their own work. Over the past year, Diode used them to produce many boards that were 95 per cent complete but unusable.

Fable 5.2 routed high-speed signals, high currents and noise-sensitive connections first. When the voice-assistant hub’s traces became tangled, it unpicked and rerouted them dozens of times. For long traces, it wrote its own routing program, which routed a trace in about a second.

Benchmark results

In Diode’s internal PCB Gym benchmark, the model places components and routes connections on boards whose original layout has been removed. Fable 5.2 scored 69 per cent and Fable 5 25 per cent. In the public BenchCAD 2.1 benchmark, the model builds 3D parts, assemblies and circuit boards from drawings and images; scores run from 0 to 100, with partial credit. At the highest reasoning effort, Fable 5.2 scored 83.5 points, Opus 5.5 65.3 and Fable 5.1 44.4.

The figures come from Anthropic and its partner. The New Stack and The Verge reported the company’s data but did not test it themselves.

Fable 5.2 is available on all platforms, and its API price is $8 per million input tokens and $40 per million output tokens. For most coding and knowledge-work tasks, the company still recommends Opus 5.5.

What this is based on

1. Claude can design electronic prototypes from start to finish

   Publisher: www.anthropic.com · published 9 November 2026 at 02:00 (UTC+02:00) · retrieved 10 November 2026 at 05:09 (UTC+02:00) · discovery channel: web

2. Introducing Claude Fable 5.2 and Claude Mythos 5.2

   Publisher: www.anthropic.com · published 9 November 2026 · retrieved 10 November 2026 at 05:08 (UTC+02:00) · discovery channel: web

3. github.com

   Publisher: github.com · published 9 November 2026 at 21:22 (UTC+02:00) · retrieved 10 November 2026 at 05:08 (UTC+02:00) · discovery channel: GitHub

4. During pre-release testing, Claude Fable 5.2 and Claude Mythos 5.2 worked on ambitious projects in many different fields. Here are a few. From a one-paragraph prompt, Fable 5.2 designed three…

   claudeai · Publisher: x.com · published 9 November 2026 at 23:06 (UTC+02:00) · retrieved 10 November 2026 at 04:21 (UTC+02:00) · discovery channel: X

5. bro just proved fermat's last theorem AS A DEMO for the fable 5.2 launch and casually told me he's got something bigger planned for his aie code talk lol anthropic u guys are wildin

   swyx · Publisher: x.com · published 10 November 2026 at 01:28 (UTC+02:00) · retrieved 10 November 2026 at 04:05 (UTC+02:00) · discovery channel: X

6. thenewstack.io

   Publisher: thenewstack.io · published 9 November 2026 at 21:08 (UTC+02:00) · retrieved 10 November 2026 at 05:08 (UTC+02:00) · discovery channel: web

7. Anthropic launches new Claude Fable and Mythos 5.2 models

   Publisher: www.theverge.com · published 9 November 2026 at 21:00 (UTC+02:00) · retrieved 10 November 2026 at 05:08 (UTC+02:00) · discovery channel: web

On 9 November, Der Spiegel reported on a measure from the ifo Institute, drawing on a story by the news agency dpa. The institute calls the measure its AI early indicator (KI-Frühindikator). It is designed to track how the spread of AI shows up in working life.

The researchers compared employment and new hires in 140 occupations with how easily AI can replace the tasks in each. Data entry and word processing top the list of the most exposed, followed by software developers, accountants, database and network technicians, and authors and journalists. The least exposed are craft and other manual occupations such as construction workers, carpenters, painters, cleaners and farmers.

In the most exposed occupations, full-time employment among 20- to 24-year-olds has fallen by 14 per cent since 2022. The researchers stress that AI alone does not explain the change. Demographic change and longer periods of study may also play a part, they say.

For people at the start of their careers, what matters is whether AI complements their tasks or automates them, said ifo economist Yuchen Guo. ‘In areas such as supporting office work or IT support, which have so far offered good entry opportunities, AI can already take over many tasks,’ Guo said.

Overall employment tells a different story. The ifo Institute said employment in AI-exposed occupations has grown by 1.8 per cent since 2022. Over the same period, employment in occupations that AI barely reaches has fallen by 6.4 per cent. These figures cover all age groups, so they cannot be compared directly with the figure for young people’s full-time work.

‘We have so far seen nothing of broad job cuts caused by AI,’ said ifo President Clemens Fuest. ‘But anyone starting out in working life feels the change first.’

Oliver Falck, one of the study’s co-authors, sees the loss of jobs for young people entering the workforce as an early warning sign in the labour market. The authors also see positive effects: AI can support people in their work and make them more productive.

What this is based on

1. Arbeitsmarkt: KI hat laut Ifo-Auswertung Einfluss auf Berufschancen von Einsteigern

   Publisher: www.spiegel.de · published 9 November 2026 at 13:20 (UTC+02:00) · retrieved 10 November 2026 at 05:46 (UTC+02:00) · discovery channel: web

Legal technology company Filevine described in a blog post on 9 November how its AI tools find the most relevant passages in documents for a language model to use. The author, Jakub Žovák, wrote that many AI applications succeed or fail precisely because of retrieval.

The starting point is RAG, or retrieval-augmented generation. An embedding model converts documents into vectors, or sequences of numbers, so that texts with similar meanings have vectors close to each other. The system fetches the texts closest to the question from a vector database, and a stronger model re-scores the candidates.

Filevine regards hybrid retrieval as the industry’s established approach. Hybrid retrieval runs an exact keyword search alongside vector search. Reciprocal rank fusion (RRF) merges the results, boosting documents that rank highly in both searches.

Hybrid retrieval is a good baseline but falls short of what the legal domain requires, the company said. The blog contains no measurements, so the superiority of its own methods remains the company’s claim.

Metadata as part of the vector

Legal documents carry metadata such as court level and citation count. Many AI systems can use metadata as a hard filter. Filevine says it supports soft filtering, which can express a preference: a document ideally from an appellate court or higher.

The simplest method is to re-rank similarity search results by metadata. Zilliz’s vector database, for example, supports this. The blog says the method fails if the search returns only trial court documents. In that case, no desired results remain.

Filevine says it can embed metadata directly into vectors with specialised models and weight it at query time. In an example in a blog footnote, text similarity and the document’s metadata value each count for half. The company cites the paper Utilizing Metadata for Better Retrieval-Augmented Generation as background to the method.

Synonyms and token-level comparison

Traditional keyword search does not recognise synonyms or homonyms and stumbles over legal terms, the blog says. As a fix, Filevine says it can use learned sparse retrieval. In this approach, a model weights words by their importance in context and adds related words missing from the text to the search. Ready-made models such as SPLADE are general-purpose, but the company has its own models fine-tuned for legal work.

A standard embedding model condenses a text chunk into a single vector by averaging the per-token vectors, losing some information. The ColBERT method keeps a separate vector for each token. For every token in the question, it finds the best-matching token in the document and sums the similarities of these matches. This typically makes retrieval more expressive, the blog says. The method also shows which part of the question matched which passage.

Multi-vector retrieval has mostly been used for re-ranking because its latency is higher than that of single-vector retrieval. Filevine says Google’s Muvera index removes the latency problem. The company says it can use multi-vector representation for both retrieval and re-ranking.

The blog ends by urging readers to try LOIS, the company’s legal AI. The free trial requires no credit card, according to the company’s website.

What this is based on

1. Advanced Legal Retrieval: Improving Vector & Keyword Search

   Publisher: filevine.com · published 9 November 2026 at 18:04 (UTC+02:00) · retrieved 10 November 2026 at 05:46 (UTC+02:00) · discovery channel: web

2. Advanced Legal Retrieval: Improving Vector & Keyword Search (Filevine Inc)

   Publisher: article.wn.com · published 9 November 2026 · retrieved 10 November 2026 at 05:47 (UTC+02:00) · discovery channel: web

London-based law firm Addleshaw Goddard (AG) announced on 9 November that it will adopt Harvey and co-develop the platform with Harvey. Harvey is an AI platform for legal work. From January, more than 3,400 lawyers and business services professionals in AG’s offices across Europe, the Middle East and Asia will use it.

Selection and co-development

AG said the choice followed a three-month assessment led by its 80-strong Innovation Group. In the assessment, 400 partners, lawyers, legal technologists and business services professionals tested leading platforms on real tasks. AG has not disclosed the results or which other platforms it tested.

Kerry Westland, the partner who leads the Innovation Group, told Artificial Lawyer that Harvey’s matter-specific workspaces (‘matter spaces’) were one of the key reasons for the choice. In the co-development work, AG is prioritising features for managing large projects and matters. Westland said AG will not use Legora alongside Harvey. The firm will keep its own AGPT system, but may in future use it more for testing than as its main system.

As an example of results so far, AG cites the Review Tables feature, which lawyers use to organise, compare and check large volumes of information. AG’s feedback helped ensure that work lawyers have already reviewed is preserved, and shaped the feature’s source checking, sharing and performance, the announcement said. Unite.AI points to a post published by Harvey’s engineers on 22 October. The post describes tables in which rows are documents and columns are questions. Each answer includes the model’s reasoning and source citations, and users can mark answers as reviewed or lock them. Unite.AI says its article was written by AI and checked by its editorial team.

AG and Harvey also plan to meet selected clients quarterly and, where needed, develop solutions jointly with them. In the announcement, Westland stressed that human judgement, supervision and accountability remain at the centre.

Who benefits from co-development

A growing number of large law firms are taking part in developing AI tools, Non-Billable reported. Latham, for example, is developing fund formation features with Harvey, and Pinsent Masons is building new features with Legora. Citing Bloomberg Law, Non-Billable said Harvey plans to offer the technology developed with Latham to other firms as well, without sharing revenue with Latham. Such arrangements raise the question of how much competitive advantage firms can keep for themselves, the site wrote. AG’s announcement does not say whether the features developed with AG will be reserved for its own use.

Fieldfisher trialled Legora for six weeks

Fieldfisher, which has around 1,400 lawyers and other staff, is adopting Legora, Artificial Lawyer and Líder Legal reported. The decision followed a six-week pilot in which 137 users in the firm’s European offices used the platform in their day-to-day work. Lawyers use Legora for legal research, document analysis and drafting, Líder Legal said.

Fieldfisher’s AI centre in Berlin is leading the roll-out. The firm said it is developing its own prompts, workflows and guidance alongside Legora to retain expertise and flexibility. Large law firms that have chosen neither Harvey nor Legora are becoming rarer, according to Artificial Lawyer.

What this is based on

1. Addleshaw Goddard and Harvey launch firmwide collaboration to shape AI-enabled legal delivery | Addleshaw Goddard LLP

   Publisher: www.addleshawgoddard.com · published 9 November 2026 · retrieved 10 November 2026 at 05:38 (UTC+02:00) · discovery channel: web

2. Addleshaw Goddard Picks Harvey, Fieldfisher Opts For Legora

   Publisher: www.artificiallawyer.com · published 9 November 2026 at 12:58 (UTC+02:00) · retrieved 10 November 2026 at 05:38 (UTC+02:00) · discovery channel: web

3. Addleshaw Goddard becomes latest firm to co-develop Harvey tools

   Publisher: www.nonbillable.co.uk · published 9 November 2026 at 16:59 (UTC+02:00) · retrieved 10 November 2026 at 05:39 (UTC+02:00) · discovery channel: web

4. Addleshaw Goddard, Harvey Partner On Firmwide AI Adoption - Law360 Pulse

   Publisher: www.law360.com · published 9 November 2026 at 23:18 (UTC+02:00) · retrieved 10 November 2026 at 05:39 (UTC+02:00) · discovery channel: web

5. Addleshaw Goddard to Deploy Harvey Firmwide and Co-Develop Legal AI

   Publisher: www.unite.ai · published 9 November 2026 at 12:34 (UTC+02:00) · retrieved 10 November 2026 at 05:39 (UTC+02:00) · discovery channel: web

6. Addleshaw Goddard skal rulle ut Harvey i hele firmaet og videreutvikle juridisk KI i fellesskap

   Publisher: www.unite.ai · published 9 November 2026 at 12:34 (UTC+02:00) · retrieved 10 November 2026 at 05:39 (UTC+02:00) · discovery channel: web

7. Fieldfisher boosts its global AI strategy with the deployment of Legora - Líder Legal

   Publisher: liderlegal.com · published 9 November 2026 at 09:00 (UTC+02:00) · retrieved 10 November 2026 at 05:39 (UTC+02:00) · discovery channel: web

The American Arbitration Association (AAA) said in a press release on Monday 9 November that its AI Arbitrator service is no longer limited to construction disputes. The service launched for construction disputes on 3 November 2025.

The service can now be used in all disputes between two businesses that are decided on documents alone, the release said. However, the news site Unite.AI reported that it does not cover cases involving complex questions of fact. Unite.AI says its article was written by AI and checked by its editors. An article on the Business Bearings site summarises the Unite.AI piece.

AI drafts, arbitrator decides

The AAA said the AI evaluates the merits of claims, gives reasons for its recommendations and prepares a draft award. A specially trained AAA arbitrator reviews the analysis, amends it as needed and issues the final award at the arbitrator’s own discretion.

Unite.AI describes the process in more detail, citing AAA material. The parties submit their positions, evidence and legal authorities. The AI summarises them, and each party can correct the summary of its own submissions. The parties do not see the AI’s analysis of the key issues or the draft award. The AAA said the tool does not make decisions on its own: the arbitrator is responsible for correcting any shortcomings and for the outcome.

Both parties must consent. Otherwise the case proceeds under the AAA’s standard arbitration process.

The AI was trained for commercial disputes on cases from more than 20 industries, including healthcare, financial services, energy and technology, the release said. Unite.AI reported that, according to the AAA, case data and AI outputs are not used to train the model, and OpenAI’s foundation models do not absorb case material.

Price and timetable

Unite.AI reported that under the amended AI-Led Arbitration Rules, effective 18 February 2026, the cost is $2,500 per claim and the same amount separately for a counterclaim. The price covers AAA administration and the arbitrator’s compensation. The parties can also agree at the outset that the award is advisory only and not binding on them.

The final award is due within five business days after the deadline for party feedback. The five-business-day limit applies only to this final stage, not to the handling of the whole dispute.

Unite.AI reported that the AAA’s FAQ page cites early results of 20–25 per cent faster resolution and cost savings of at least 35 per cent. The figures apply only to documents-only construction disputes, so they say nothing about how the new commercial cases will go. Nor does the Unite.AI article say what process the figures are compared against.

What this is based on

1. AAA Expands AI Arbitrator to Documents-Only Commercial Cases

   Publisher: www.unite.ai · published 9 November 2026 at 16:05 (UTC+02:00) · retrieved 10 November 2026 at 05:39 (UTC+02:00) · discovery channel: web

2. The AI Arbitrator from the American Arbitration Association®, Now Available for Commercial Cases

   Publisher: www.prnewswire.com · published 9 November 2026 at 15:30 (UTC+02:00) · retrieved 10 November 2026 at 05:39 (UTC+02:00) · discovery channel: web

3. AAA Expands AI Arbitration to Commercial Disputes

   Publisher: businessbearings.com · published 10 November 2026 at 02:09 (UTC+02:00) · retrieved 10 November 2026 at 05:39 (UTC+02:00) · discovery channel: web

Northern Trust and Microsoft announced the agent in a joint press release on Monday 9 November. The Chicago-based bank uses it when onboarding new clients to its Front Office Solutions service, which produces performance and analytics reports for clients such as foundations, pension funds and large family offices.

Pete Cherecwich, the bank’s chief operating officer, told CIO.com that a large institutional client can have huge volumes of data, particularly from alternative investments. Several years of transactions and cash flows can reach the bank in widely differing formats. The bank previously converted the data by hand before moving it into its reporting environment, CIO.com reported.

Rule library and human review

The bank’s financial experts and Microsoft engineers built a rule library for the agent, Microsoft’s news site reported. The library lets the agent recognise the same item even when clients’ workbooks give it different names. Cherecwich cited capital calls and valuations as examples. The team gave the model the industry’s vocabulary and ran it on anonymised data from several past clients.

The engineers came from Microsoft Frontier Company, which places its developers in customers’ teams. Microsoft launched the $2.5 billion unit in July, CIO.com reported. Microsoft used only anonymised data to improve the agent, and client data stayed under the bank’s control, the press release said.

The agent runs on Microsoft’s Azure cloud and uses the Microsoft Foundry AI platform. Analysts review and approve its output in Excel and SharePoint. ‘The agent does a lot of the heavy lifting, and then our people review the output and handle the things where there’s uncertainty,’ Cherecwich told CIO.com.

What the figures measure

The bank measured the time saving in real client cases after deploying the agent. The measurement covered data conversion, analyst review, exception handling and final validation. It compared the previous workflow’s baseline of about 2.5 weeks with the average duration of cases handled with the agent. The time fell by 83 per cent, the press release said. The number of cases measured was not disclosed.

Microsoft wrote on X that client onboarding fell from 2.5 weeks to three days. The press release and its footnote, however, limit the figure to data preparation. After preparation, the bank carries out its regular reconciliation before the client moves onto the service, Microsoft’s article said. The reconciliation compares the bank’s books and records with those of the previous administrator.

The quality figure comes from a separate measurement. The agent achieved an F1 score of 96.5 per cent in a controlled test of seven cases drawn from real but anonymised client workbooks. Its first, uncorrected output was compared with answers prepared by analysts, which the agent had not seen. F1 combines two things: how much of the data the agent produces is correct, and how much of the correct data the agent finds.

Microsoft’s article describes the remaining 3.5 per cent as data the agent is unsure how to classify, which is routed to humans for review. The F1 score does not, however, directly show what share of the data ends up with humans.

Cherecwich said in Microsoft’s article that the bank does not plan to let AI make $100 million payments on its own any time soon. The bank is also exploring the use of agents in its other products, the article said.

What this is based on

1. Northern Trust and Microsoft Frontier Company collaborate to transform client onboarding through AI

   Publisher: news.microsoft.com · published 9 November 2026 at 17:58 (UTC+02:00) · retrieved 10 November 2026 at 05:40 (UTC+02:00) · discovery channel: web

2. Northern Trust and Microsoft Frontier Company accelerate client onboarding with AI

   Publisher: msft.it · published 9 November 2026 at 09:57 (UTC+02:00) · retrieved 10 November 2026 at 05:40 (UTC+02:00) · discovery channel: web

3. Parsing complex financial data doesn't have to take forever. Northern Trust and Microsoft Frontier Company cut client onboarding from 2.5 weeks to 3 days, with human expertise still at the…

   Microsoft · Publisher: x.com · published 9 November 2026 at 21:16 (UTC+02:00) · retrieved 10 November 2026 at 04:20 (UTC+02:00) · discovery channel: X

4. Northern Trust streamlines client onboarding with AI agent

   Publisher: www.cio.com · published 9 November 2026 · retrieved 10 November 2026 at 05:40 (UTC+02:00) · discovery channel: web

5. Microsoft Corporation (MSFT) Stock: Jumps as Northern Trust Unveils AI-Powered Onboarding Agent

   Publisher: blockonomi.com · published 9 November 2026 at 20:37 (UTC+02:00) · retrieved 10 November 2026 at 05:40 (UTC+02:00) · discovery channel: web

BlackLine, which makes finance software, issued a press release on 9 November 2026 about an AI roadmap it is carrying out with SAP. The roadmap covers the record-to-report process: the accounting chain from recording business transactions to financial reporting.

BlackLine plans to integrate its Verity AI agents with SAP’s Joule and run them on SAP Business AI Platform, the release said. An AI agent is a program that carries out tasks using an AI model. This is a plan. The release gives neither a launch date nor a price for the agents. The descriptions of what the agents do are the company’s own.

The first phase is due to bring three agents to period-end close, when the books are reconciled and finalised before reporting:

• Verity Prepare analyses subledger activity, compiles supporting documentation and drafts balance sheet reconciliations.
• Verity Match matches large volumes of transactions in environments with several enterprise resource planning (ERP) systems.
• Verity Accruals reviews data from ERP systems and other sources, calculates period-end accruals and proposes journal entries for them. The entries are approved in BlackLine, from which they could be posted to several ERP systems.

BlackLine said Verity Accruals complements SAP’s own Accruals Agent, part of SAP Financial Closing Assistant. BlackLine’s agent is intended for complex journal entries, for example those that draw on data from several SAP and third-party sources or are posted to several ERP systems.

Jeremy Ung, BlackLine’s chief technology officer, explained the partnership by saying that the built-in AI of ERP systems automates transaction-level tasks within the system. Period-end close, however, requires data integrity across the entire accounting chain, Ung said. In the release, Catharina Riedl, global head of sales at SAP Finance, called the collaboration the next milestone in the partnership.

BlackLine is an SAP platinum partner and part of the SAP Solution Extensions programme. The company said its solutions in the programme have passed the programme’s quality assessment and SAP support services are available for them.

BlackLine promises more information at blackline.com/sap. BlackLine and SAP will hold a webinar on 3 December 2026 on SAP Autonomous Finance and agentic financial operations.

What this is based on

1. BlackLine Announces AI Roadmap in Alignment with SAP Partnership to Deliver AI-Powered Gold Standard for Record-to-Report

   Publisher: www.globenewswire.com · published 9 November 2026 at 16:00 (UTC+02:00) · retrieved 10 November 2026 at 05:41 (UTC+02:00) · discovery channel: web

Reuters reported on the evening of Monday 9 November that it had seen emails Anthropic sent to bankers. The Economic Times, investingLive and Investing.com are among the outlets that have carried the report.

The first email included a copy of the IPO filing as an attachment, Reuters said. According to the filing, Anthropic planned to sell about 867 million Class A shares at $70 to $80 each. At the top of the range, the offering would have raised $69.3 billion. In a follow-up email, the company said the pricing details were inaccurate and promised to update the filing later. Anthropic did not comment to Reuters.

The company did not say which figures were wrong, the Turkish technology site ShiftDelete.Net reported. The retracted figures had already spread: China’s 36Kr published them on Tuesday, citing Sina Finance, without mentioning the retraction.

A company going public files an official registration statement, the S-1, with the US Securities and Exchange Commission (SEC). ShiftDelete.Net said the S-1 that Anthropic filed on 26 October and its 2 November amendment both leave the price range and share count blank. The filings confirm a planned Nasdaq listing under the ticker ANT.

The $26 billion investment is also uncertain

Reports based on the same copy of the filing said Google, Nvidia, AMD, Broadcom, Micron, Microsoft and TSMC would invest a total of $26 billion in Anthropic. They would buy shares directly from Anthropic at the IPO price. ShiftDelete.Net cites CNBC as the source for this. Anthropic has not confirmed the arrangement. investingLive said it is unclear whether the correction also applies to it.

Investors want a lower valuation

Some institutional investors want a valuation below the $1.5 trillion target discussed as recently as last week, The Wall Street Journal reported. That figure was itself considerably below the $2 trillion floated a few months ago. TechCrunch carried the report. Bloomberg reported on 6 November that some prospective buyers had discussed a valuation of $1.1 trillion to $1.3 trillion, ShiftDelete.Net said. A sell-off in AI stocks wiped about $2.5 trillion off the market value of companies in the Nasdaq Composite index last week, investingLive said.

The retraction is open to different readings. investingLive considers the errors procedural rather than a sign of a change of plan. It added, however, that the final terms are now more likely to differ from the figures that circulated. ShiftDelete.Net considers it unusual for the pricing details of such a large offering to be sent and retracted on the same evening. In its view, the terms are still being negotiated.

Timetable

Anthropic plans to set a valuation range this week, and trading is due to begin on Thursday or Friday next week at the earliest, the WSJ reported. ShiftDelete.Net notes that this report was published before the retraction. Nothing so far suggests the timetable will change, however, investingLive said. The Defiant reported on Monday, citing IFR, that Anthropic would launch its IPO on Tuesday.

What this is based on

1. Anthropic拟通过首次公开募股筹集至多693亿美元-36氪

   Publisher: 36kr.com · published 10 November 2026 · retrieved 10 November 2026 at 05:17 (UTC+02:00) · discovery channel: web

2. Anthropic to Launch Nasdaq IPO Tuesday, IFR Reports | The Defiant

   Publisher: thedefiant.io · published 9 November 2026 at 22:24 (UTC+02:00) · retrieved 10 November 2026 at 05:17 (UTC+02:00) · discovery channel: web

3. Anthropic reportedly plans to move forward with IPO plans | TechCrunch

   Publisher: techcrunch.com · published 10 November 2026 at 03:27 (UTC+02:00) · retrieved 10 November 2026 at 05:17 (UTC+02:00) · discovery channel: web

4. Anthropic sends out, then retracts details of IPO filing, emails show

   Publisher: economictimes.indiatimes.com · published 10 November 2026 at 04:18 (UTC+02:00) · retrieved 10 November 2026 at 05:17 (UTC+02:00) · discovery channel: web

5. Anthropic IPO Price Range Retracted After $69.3B Report

   Publisher: shiftdelete.net · published 10 November 2026 at 04:18 (UTC+02:00) · retrieved 10 November 2026 at 05:17 (UTC+02:00) · discovery channel: web

6. Anthropic retracts IPO pricing details sent to bankers, says figures were inaccurate

   Publisher: investinglive.com · published 10 November 2026 at 04:28 (UTC+02:00) · retrieved 10 November 2026 at 05:17 (UTC+02:00) · discovery channel: web

7. Anthropic retracts IPO details after sending filing to bankers - Reuters

   Publisher: sg.finance.yahoo.com · published 10 November 2026 at 05:08 (UTC+02:00) · retrieved 10 November 2026 at 05:18 (UTC+02:00) · discovery channel: web

Trump disclosed the FBI investigation after speaking to Huang on Monday, Helsingin Sanomat reported on Tuesday, citing Reuters and the BBC. According to Trump, Huang said he had not known the woman was a spy. Trump said he believed him. ‘I think he will come through this fine,’ Trump said, HS reported.

The woman is Alice Huang, a former executive of Megaspeed, a Singapore-based data centre company. In a message to staff on Sunday, Jensen Huang admitted being unfaithful to his wife and apologised.

Nvidia’s board has launched an investigation led by its independent members. The board said it had no reason to believe that national security or Nvidia’s secrets had been compromised. Huang will remain chief executive during the investigation. Nvidia’s shares fell to a seven-month low on Monday, HS reported.

Suspicions that chips ended up in China

US authorities suspect that Nvidia AI chips whose export to China is restricted reached the country via Megaspeed. According to trade data cited by The New York Times, Megaspeed’s Malaysian sister companies bought at least $4.8 billion of Nvidia equipment, including about 136,000 GPUs, between June 2024 and February 2026 (Asia Times).

Bloomberg reported in late 2025 that Megaspeed had bought about $4.6 billion of Nvidia chips since its founding, far more than its own data centres could hold. Bloomberg could not, however, show that the chips had ended up in China (Next). Megaspeed said reviews by the US Commerce Department, Malaysian authorities and Nvidia had found no violations.

Intelligence officials believe Chinese intelligence recruited Alice Huang to gather information about Jensen Huang and to influence Nvidia’s efforts to loosen export controls, TechCrunch reported. China’s foreign ministry has called the allegations slander. TechCrunch notes that Jensen Huang had been pushing for sales to China even before he and Alice Huang reportedly met in early 2024.

As recently as late October, Huang told Bloomberg that their relationship and dealings had been related to Megaspeed and entirely ordinary. Shakeel Hashim of the Transformer newsletter has called for Huang to resign, arguing that Huang misled staff, investors and the public. Hashim also writes that Nvidia’s general counsel, Tim Teter, reportedly contacted the Commerce Department about its Megaspeed investigation, leading the department to transfer the case to the Justice Department. Hashim does not name the source of this information.

Congress demands an explanation

In a letter dated 4 November, the Republican chairs of three House committees asked Huang to attend a closed-door briefing on 18 November, Asia Times reported. The Democrats won both chambers of Congress in the midterm elections. The scandal weakens Nvidia’s negotiating position as the new Congress takes over committee leadership in January, Bloomberg Intelligence said (The Decoder).

Analysts believe the Remote Access Security Act could now advance quickly, Asia Times reported. The bill would extend export controls from chips to cloud computing and stop Chinese companies from remotely renting servers with advanced Nvidia chips in third countries such as Malaysia. Passed by the House in January, it awaits consideration by the Senate.

What this is based on

1. Vakoilu | Trump: FBI tutkii Nvidian johtajan suhdetta epäiltyyn kiinalaisvakoojaan

   Publisher: www.hs.fi · published 10 November 2026 at 01:57 (UTC+02:00) · retrieved 10 November 2026 at 05:19 (UTC+02:00) · discovery channel: web

2. Nvidia espionage scandal weakens Huang's lobby against tighter AI chip export controls

   Publisher: the-decoder.com · published 9 November 2026 at 13:38 (UTC+02:00) · retrieved 10 November 2026 at 05:20 (UTC+02:00) · discovery channel: web

3. Jensen Huang’s affair with alleged Chinese spy puts Nvidia’s China ambitions under scrutiny | TechCrunch

   Publisher: techcrunch.com · published 9 November 2026 at 20:09 (UTC+02:00) · retrieved 10 November 2026 at 05:20 (UTC+02:00) · discovery channel: web

4. Nvidia’s Huang spy scandal signals tighter US chip controls - Asia Times

   Publisher: asiatimes.com · published 9 November 2026 at 23:02 (UTC+02:00) · retrieved 10 November 2026 at 05:20 (UTC+02:00) · discovery channel: web

5. Jensen Huang should resign

   Shakeel Hashim · Publisher: www.transformernews.ai · published 9 November 2026 at 20:59 (UTC+02:00) · retrieved 10 November 2026 at 04:15 (UTC+02:00) · discovery channel: RSS feed

6. Espionnage chinois : NVIDIA enquête sur son propre CEO

   Alexandre Laurent · Publisher: next.ink · published 9 November 2026 at 12:03 (UTC+02:00) · retrieved 10 November 2026 at 04:26 (UTC+02:00) · discovery channel: RSS feed

7. Why would anyone trust Nvidia's "independent investigation" into Jensen Huang following news that the company apparently intervened to intimidate Commerce Dept. officials who were conducting their…

   CaseyNewton · Publisher: x.com · published 9 November 2026 at 19:02 (UTC+02:00) · retrieved 10 November 2026 at 04:10 (UTC+02:00) · discovery channel: X

Google announced on Monday 9 November two new systems for its eighth-generation TPU line-up. The TPU is Google’s own chip for AI computing. TPU 8tx and TPU 8i Max are not new chips, MarkTechPost and The Decoder reported. Google uses the TPU 8t and TPU 8i chips it unveiled in April and changes their packaging, host processors, memory and networking.

Google says this is the first time it has released four TPU systems in the same year.

TPU 8tx protects data during processing

Data is usually encrypted at rest and in transit. When a processor works on the data, it is decrypted in memory. Google says TPU 8tx keeps data and model weights protected and isolated while in use as well. The system is aimed at regulated industries and government.

TPU 8tx pairs the TPU 8t chip with 5th Gen AMD EPYC processors and uses the Boardfly network from TPU 8i. Google said it cuts latency by up to half compared with TPU 8t, and that the protection barely slows computation. Customers can fine-tune and distil models and train them with reinforcement learning on the same system, the company said. Data does not have to move from one cluster to another. The Decoder noted that Nvidia already offers similar confidential computing for its GPUs.

TPU 8i Max combines two chips

TPU 8i Max puts two TPU 8i chips in one package. Compared with TPU 8i, it doubles compute, on-chip SRAM and HBM bandwidth, Google said. HBM is fast memory packaged next to the chip. Twice as many chips fit in the same rack space. Google is offering the system for AI agents that work through entire codebases or hours of video.

MarkTechPost doubled TPU 8i’s published figures and arrived at about 20.2 petaflops at FP4 precision and about 17.2 TB/s of memory bandwidth per package. Google did not publish these figures or the amount of HBM, although The Decoder writes that memory doubles.

Google gave no availability dates or prices. MarkTechPost cites a Bloomberg report that both systems will become available in 2027.

TPU Sync speeds up cache transfers

Google also released TPU Sync, an open-source C++ library. A language model stores intermediate results for text it has already processed in a KV cache so that it does not have to recompute them. TPU Sync moves this cache between chip memory, server main memory and other TPU nodes.

MarkTechPost summarised Google’s own measurements. In them, median cache transfer latency for the Qwen3 235B model fell from several seconds to about 230 milliseconds. First-token latency fell by an average of 60 per cent on the 235B model and 90 per cent on the 480B model when there were few concurrent requests. The measurements were run on Ironwood chips (TPU v7x), not on the new systems.

Its GitHub page says TPU Sync is tested and optimised for TPU v7x and v6e chips and that its APIs are still changing quickly. PyPI has only development releases so far. The project’s former package names, tpu-raiden-jax and tpu-raiden-torch, are not under its control and should not be installed.

What this is based on

1. github.com

   Publisher: github.com · source publication time unknown · retrieved 10 November 2026 at 05:28 (UTC+02:00) · discovery channel: GitHub

2. Google Unveils TPU 8i Max and TPU 8tx: A Dual-Chip TPU and a Confidential TPU

   Publisher: www.marktechpost.com · published 9 November 2026 at 19:09 (UTC+02:00) · retrieved 10 November 2026 at 05:27 (UTC+02:00) · discovery channel: web

3. Google ships four TPU variants in one year as it ramps up competition with Nvidia

   Publisher: the-decoder.com · published 9 November 2026 at 14:59 (UTC+02:00) · retrieved 10 November 2026 at 05:27 (UTC+02:00) · discovery channel: web

4. Google、第8世代TPUに「8tx」と「8i Max」を追加 機密データ処理と大規模AIエージェントに特化

   Publisher: www.itmedia.co.jp · published 10 November 2026 at 09:26 (UTC+02:00) · retrieved 10 November 2026 at 05:27 (UTC+02:00) · discovery channel: web

5. Google expands AI chip lineup as Nvidia faces CEO spy scandal

   Publisher: au.investing.com · published 9 November 2026 at 14:46 (UTC+02:00) · retrieved 10 November 2026 at 05:28 (UTC+02:00) · discovery channel: web

6. Google’s chip business is on a faster track than Nvidia’s own AI boom

   Publisher: finance.yahoo.com · published 9 November 2026 at 23:17 (UTC+02:00) · retrieved 10 November 2026 at 05:28 (UTC+02:00) · discovery channel: web

7. TorchTPU is now open source: Run native PyTorch on TPUs | Google Cloud Blog

   Publisher: cloud.google.com · published 20 October 2026 · retrieved 10 November 2026 at 05:28 (UTC+02:00) · discovery channel: web

8. Google Escalates Nvidia Rivalry With New Versions of Its TPU Systems

   Publisher: www.bloomberg.com · published 9 November 2026 at 14:30 (UTC+02:00) · retrieved 10 November 2026 at 05:27 (UTC+02:00) · discovery channel: web

The British business newspaper the Financial Times (FT) reported the deal on Sunday 8 November, citing a person familiar with the matter. Other outlets have summarised the FT’s report. In a story published by CNBC, Reuters said it could not immediately verify the report. Anthropic and OnePWR did not respond to Reuters’ requests for comment.

Anthropic agreed the deal in late September, the FT reported. The deal is worth about $3.4 billion, and the generators have a combined capacity of up to 1.7 gigawatts.

Anthropic took over OnePWR’s orders

OnePWR Solutions, an equipment supplier based in San Antonio, Texas, ordered the generators from Caterpillar about two years ago. DCD reported that OnePWR placed the order before recent price increases but needed financing to complete the transaction. In the end, Anthropic took over the Caterpillar orders itself, although it was originally meant to be a OnePWR customer.

Anthropic did not buy OnePWR, and none of OnePWR’s employees moved to Anthropic. The deal does, however, include the engineering and construction work needed to install the generators. DCD said the generators are expected to be available within two years.

TechCrunch asks why OnePWR failed to secure financing despite Anthropic’s interest. The publication speculates that Anthropic wanted to keep the generating capacity for itself before a competitor could buy it.

Power has become a bottleneck

The deal is the largest known single purchase of power-generation equipment by an AI company, the FT reported, as summarised by Heise. The paper said AI companies are using their financial muscle to bypass intermediaries because delivery times for gas turbines and generators have stretched to years.

Tom’s Hardware wrote that turbine orders worldwide reportedly exceed 100 gigawatts, against an estimated manufacturing capacity of 60–70 gigawatts. The average delivery time in the US is six years, it said. TechCrunch reported that the backlog of gas turbine orders runs into the early 2030s, and connecting to the grid can take just as long.

TechCrunch notes that Anthropic has usually leased capacity rather than buying it. In its S-1 filing in May, the company disclosed that it is paying SpaceX subsidiary xAI $1.25 billion a month for computing power until May 2029. TechCrunch sees the generator deal as a last resort. The FT reported that Anthropic has compute commitments worth nearly $900 billion running to 2045.

The deal has also drawn criticism. Tom’s Hardware reported that Shakeel Hashim, founder of ReadTransformer, wrote on X: ‘This is pretty grim—these kinds of generators are terrible for the environment.’ TechCrunch also points out that gas exposes companies to price swings. One analyst estimates that gas prices could more than triple in some markets by the end of the decade.

What this is based on

1. Anthropic signs $3.4bn gas generator deal to support 1.7GW of data center power

   Publisher: www.datacenterdynamics.com · published 9 November 2026 at 15:33 (UTC+02:00) · retrieved 10 November 2026 at 05:28 (UTC+02:00) · discovery channel: web

2. Anthropic skips six-year waitlist with massive $3.4 billion gas turbine generator deal — firm takes over OnePWR's orders, secures 1.7GW of capacity to power AI data centers

   Publisher: www.tomshardware.com · published 9 November 2026 at 15:45 (UTC+02:00) · retrieved 10 November 2026 at 05:28 (UTC+02:00) · discovery channel: web

3. 앤트로픽, 데이터센터 전력 확보에 4.6조 투입…가스 발전기 직접 구매 - AI타임스

   Publisher: www.aitimes.com · published 9 November 2026 at 05:35 (UTC+02:00) · retrieved 10 November 2026 at 05:29 (UTC+02:00) · discovery channel: web

4. Milliarden-Deal für Gasturbinen: Anthropic sichert sich Energiekapazitäten

   Publisher: www.heise.de · published 9 November 2026 at 22:13 (UTC+02:00) · retrieved 10 November 2026 at 05:29 (UTC+02:00) · discovery channel: web

5. Anthropic joins AI’s natural gas frenzy, buys $3.4B worth of turbines | TechCrunch

   Publisher: techcrunch.com · published 9 November 2026 at 18:56 (UTC+02:00) · retrieved 10 November 2026 at 05:29 (UTC+02:00) · discovery channel: web

6. Anthropic récupère des générateurs d’électricité au gaz pour 3,4 milliards de dollars

   Publisher: next.ink · published 9 November 2026 at 16:39 (UTC+02:00) · retrieved 10 November 2026 at 05:29 (UTC+02:00) · discovery channel: web

7. Anthropic secures $3.4 billion generator deal for AI data centers, FT reports

   Publisher: www.cnbc.com · published 8 November 2026 at 15:50 (UTC+02:00) · retrieved 10 November 2026 at 05:30 (UTC+02:00) · discovery channel: web

GitHub user thisisme6598 published a report in November saying that Google has abandoned a privacy commitment on Gemini. The report scored 444 points on Hacker News.

Google introduced Workspace extensions to its Gemini assistant in 2023 and promised that human reviewers would not see personal content from Gmail, Docs and Drive, the report says. Help page 16836988 and a 2023 blog post that Google edited on 16 October now state that people may review connected emails and files, the author wrote. In guidance the author quotes, Google advises: ‘don’t connect your apps to Gemini Apps if they contain confidential information that you wouldn’t want a reviewer to see.’

Gemini searches without being asked

The report concerns the consumer version of Gemini, where the Connected Apps setting lets Gemini retrieve information from the user’s Gmail and Drive. In the author’s tests, Gemini searched in the background when the conversation mentioned an everyday topic such as a trip, a bill or a reservation. The user did not need to tag the request with @Gmail or @Drive.

Results for the same question varied because a large language model decides on tool calls, in this case searches for information, probabilistically. The author identifies four outcomes: no search, a general summary, a direct quotation of a few sentences that often contains personal data, and the full text of a message or file.

The author told Gemini about wondering whether the broadband subscription cost too much. On different occasions, Gemini pulled details such as a name, a customer account number and a payment confirmation number from an email from the internet provider into its answer. Sometimes it reproduced the entire message, headers included. Because results vary from run to run, a single test without personal data does not disprove the finding, the author said.

Tests rely on the model’s own output

The findings rest on one person’s experiments and have not been independently verified. The author recommends running the tests with the Flash model at High Thinking reasoning effort or a higher setting. The author established what the search retrieved by asking Gemini to print the text the search tool returned.

The tests show what content enters the conversation. The claim that conversations may be passed to human reviewers is based on Google’s guidance, not on any observation of review.

The author also argues that Google’s automated tools for removing personal data do not adequately recognise items such as account numbers, confirmation numbers and combinations of names, places and events. The author cites Google’s own documentation.

PithWire, daily.dev and aigovernance.com are among the outlets that have covered the report. They rely on the GitHub report and do not say they carried out tests of their own. The site aigovernance.com extends the issue to companies’ Workspace data, although the report deals with the consumer version. Organisations subject to GDPR may need to reassess the legal basis for processing the data, the site says. Google has not commented publicly on the change, PithWire reported.

What this is based on

1. github.com

   Publisher: github.com · source publication time unknown · retrieved 10 November 2026 at 05:35 (UTC+02:00) · discovery channel: GitHub

2. Google Reversed Its Gemini Privacy Commitment. Enterprise…

   Publisher: aigovernance.com · published 9 November 2026 at 23:05 (UTC+02:00) · retrieved 10 November 2026 at 05:35 (UTC+02:00) · discovery channel: web

3. Google Gemini sends personal files to human reviewers | PithWire

   Publisher: pithwire.com · published 10 November 2026 at 02:07 (UTC+02:00) · retrieved 10 November 2026 at 05:35 (UTC+02:00) · discovery channel: web

4. Google Gemini routes personal Drive files and emails to human annotators | daily.dev

   Publisher: daily.dev · published 9 November 2026 at 23:15 (UTC+02:00) · retrieved 10 November 2026 at 05:35 (UTC+02:00) · discovery channel: web

US President Donald Trump set out his position on AI regulation on CBS’s 60 Minutes, broadcast on Sunday 8 November. The interview was recorded at the White House on Friday. It was Trump’s first sit-down interview since the Republicans lost both the House of Representatives and the Senate in the midterm elections.

Interviewer Ross Douthat asked whether Trump would work with the Democrats on AI safety. Trump said he would. Guardrails are limits placed on the development and use of AI.

‘I don’t mind guardrails. But I don’t wanna put us in a deep second position. We’re leading number one. And China has no guardrails, by the way,’ Trump said. In China, he said, whatever President Xi Jinping says goes.

Trump says the risks are manageable

Douthat pointed to AI industry leaders who put the chance of AI destroying humanity at, for example, 10 per cent. Trump played down the fear: ‘I think that the risk is not nearly as real as it-- as they talk. And I think the risk can be managed.’

He gave competition with China as his reason. If the US does not have AI, China will, and the US will then be in an incredibly bad position. Everything, he said, is between the US and China; there is no one else.

Business news site CFO Dive reported that Trump did not say which safeguards he would support or outline a regulatory framework. Reuters noted that Trump has previously wanted technology companies to police themselves and said it is unclear how much regulation he will now accept. As recently as 15 September, Trump called warnings about AI’s dangers a hoax on Truth Social, CFO Dive reported.

AI regulation featured prominently in the election campaigns, CFO Dive said, and in some races candidates talked about data centre construction, electricity prices and the effects on jobs. Morningstar analyst Brian Colello wrote that the previous week’s fall in AI stocks such as Nvidia had been attributed mainly to the Democrats’ victory and increased regulatory risk.

Claim of threat against an AI executive

The Verge reported that Trump said in the interview that the Justice Department manages the risk. He said the administration was told that an executive at a major AI company had done something bad with AI. ‘We called them very strongly and said, “Get ready to go to jail. Please give your wife a kiss good night.”’

Trump did not name the person or the company. He said the executive then changed course completely. The claim is Trump’s own and unspecified. The passage is missing from the transcript CBS published online, so the information rests on The Verge’s report.

What this is based on

1. 60 Minutes Transcript: Donald Trump, the 2026 post-midterm elections interview

   Publisher: www.cbsnews.com · published 9 November 2026 at 02:00 (UTC+02:00) · retrieved 10 November 2026 at 05:36 (UTC+02:00) · discovery channel: web

2. Trump’s 60 Minutes interview touched on AI safety, threat to arrest an AI exec

   Publisher: www.theverge.com · published 9 November 2026 at 17:50 (UTC+02:00) · retrieved 10 November 2026 at 05:36 (UTC+02:00) · discovery channel: web

3. Trump signals openness to AI guardrails after ‘blue tsunami’

   Publisher: www.cfodive.com · published 9 November 2026 · retrieved 10 November 2026 at 05:36 (UTC+02:00) · discovery channel: web

4. Trump says AI risks can be managed, signals openness to guardrails

   Publisher: ca.finance.yahoo.com · published 9 November 2026 at 02:33 (UTC+02:00) · retrieved 10 November 2026 at 05:36 (UTC+02:00) · discovery channel: web

5. Trump says AI risks can be managed, open to guardrails

   Publisher: www.siliconreport.com · published 9 November 2026 at 07:44 (UTC+02:00) · retrieved 10 November 2026 at 05:36 (UTC+02:00) · discovery channel: web

6. US President Donald Trump says AI risks can be managed, signals openness to guardrails

   Publisher: indianexpress.com · published 9 November 2026 at 06:26 (UTC+02:00) · retrieved 10 November 2026 at 05:36 (UTC+02:00) · discovery channel: web

James Tamim, who edits a Digital Fairness Act newsletter, reported on Monday 9 November that Contexte, a French outlet covering EU policy, had published a leaked draft of the European Commission’s Digital Fairness Act. Contexte journalist Marie Colin published the 71-page draft, Tamim wrote. Euronews and Bloomberg have also seen the text, and Argentina’s Infobae relayed Bloomberg’s report. A Commission spokesperson told Euronews that the Commission does not comment on leaks.

Right to switch off addictive features

All users would gain the right to switch off autoplay, notifications unrelated to their own activity, streak-style rewards for consecutive use, and chatbot features that simulate relationships to keep people in the app, Euronews reported.

Bloomberg reported, via Infobae, that companies would have to let adults switch off features that encourage compulsive or excessive use. The draft also covers AI systems that simulate personal relationships to keep people talking, Bloomberg said.

Infobae noted that the EU already tackles addictive design through the Digital Services Act (DSA), but only for very large platforms. The new law would also reach app developers, game studios and streaming services, Infobae said.

Euronews reported that the draft is the first EU consumer-law text to regulate AI agents that make purchases on a user’s behalf. Users would choose which purchases they must approve themselves, and an agent would have to disclose when a suggestion is sponsored.

Banned design practices and fines

According to Euronews, the draft lists ten manipulative design practices that would always be deemed unfair, so authorities would not need to prove harm case by case. They include fake notices that someone has contacted the user and subscriptions that are disproportionately harder to cancel than to take out. The Commission could add new practices to the list without lawmakers’ approval.

Minors would be barred from paid loot boxes and from game advantages bought with money. A free trial would turn into a paid subscription only if the consumer explicitly agrees.

The Commission could fine a company up to 6 per cent of its global turnover, Euronews reported. Tamim said the fine is tied to the proposed reform of consumer protection cooperation and would apply to cases the Commission investigates itself.

Proposal needs approval from Parliament and member states

The Commission is due to present the proposal on Wednesday 11 November. The regulation would replace the 2005 Unfair Commercial Practices Directive and apply directly in all 27 member states. Its rules would apply 12 months after it enters into force, but the European Parliament and member states must first approve it.

Germany already opposes the proposal. Euronews cited an Agence Europe report on a German government non-paper dated 14 October. Germany argues that the EU should amend existing consumer directives rather than create a new horizontal law, and that the switch from a directive to a regulation has not yet been sufficiently justified.

What this is based on

1. Contexte publishes leaked draft of Digital Fairness Act

   Publisher: newsletter.digitalfairnessact.com · published 9 November 2026 at 16:09 (UTC+02:00) · retrieved 10 November 2026 at 05:37 (UTC+02:00) · discovery channel: web

2. EU to blacklist ten online design 'traps', leaked draft shows

   Publisher: www.euronews.com · published 9 November 2026 at 17:05 (UTC+02:00) · retrieved 10 November 2026 at 05:38 (UTC+02:00) · discovery channel: web

3. Unión Europea quiere que los usuarios puedan apagar la reproducción automática y las rachas de las apps

   Publisher: www.infobae.com · published 9 November 2026 at 20:03 (UTC+02:00) · retrieved 10 November 2026 at 05:38 (UTC+02:00) · discovery channel: web

4. EU Plans Crackdown on Addictive Social Media and AI Bot Features

   Publisher: www.bloomberg.com · source publication time unknown · retrieved 10 November 2026 at 05:37 (UTC+02:00) · discovery channel: web

Statnett announced the decision on Monday 9 November, and the Norwegian government presented a plan for a data centre licensing regime the same day.

What the pause covers

Enerdata and Baxtel, both summarising Statnett’s press release, reported that the pause halts new capacity reservations. Statnett is also suspending the processing of new applications that seek approval of a data centre’s functional properties. Pending applications for facilities that would connect directly to the 420-kilovolt transmission grid are also on hold. Baxtel reported that law firm Wiersholm believes the restriction covers connections of more than 100 megawatts, for which Statnett handles the applications.

Statnett said this is not a general halt on connections. Existing connections remain. Facilities whose functional requirements have already been approved can still connect. Statnett will process applications it has already received for facilities due to start operating soon on the regional grid. It may also assess some applications for smaller facilities if they are due to connect before the pause ends.

Statnett said the pause reduces risks to the power system. During the pause, the company will draw up requirements for data centres on staying connected during faults, the permitted rate of load change and reducing the risk of grid oscillation (Baxtel). The requirements will go out for consultation. The energy regulator RME must approve them before the pause is lifted.

Reuters (via The Economic Times), by contrast, links the pause to the government’s licensing regime: no reservations will be made until the regime is in place. Baxtel reported that the government plans to present the regime’s main principles only in autumn 2027, after the August deadline.

Connection requests exceed the country’s peak demand

Statnett said Norway has about 800 megawatts of data centre capacity connected and 4.2 gigawatts reserved. A further 9.1 gigawatts is in the queue, and requests have been made for another 18 gigawatts. Norway’s peak electricity consumption is just over 25 gigawatts. DCD puts the additional capacity at 31.3 gigawatts in total but notes that not all the projects would be built.

Energy Minister Terje Aasland said Statnett has received connection applications for 26,000 megawatts this year, 75 per cent of them from data centres (Reuters). Norway produces more electricity than it consumes. Reuters reported, however, that forecasts suggest data centres could cause a power deficit and push up prices.

Licensing regime would favour critical services

‘We want data centers in Norway, but we can’t have all types of data centers,’ said Digitalisation Minister Torgeir Micaelsen. Micaelsen said the regime would favour facilities serving defence, the police, emergency preparedness, healthcare, mobile networks and the financial system. Global cloud companies and cryptocurrency firms would have restricted access to land and electricity (Reuters). Capacity, as cited by Baxtel, reported that AI computing for Norwegian, Nordic and European needs would also be considered beneficial. The government said the regime would also cover existing data centres, with transitional arrangements.

The government wants new data centres to pay the grid costs they cause. Data centres may also be required to build their own power generation, such as gas power plants with carbon capture.

Briefs reported that Reynir Johannesson, head of the industry association Norsk Datasenterindustri, said the decision creates uncertainty for investments worth several billion kroner a month.

What this is based on

1. Norway halts grid connection applications for new data centers

   Publisher: www.datacenterdynamics.com · published 9 November 2026 at 19:48 (UTC+02:00) · retrieved 10 November 2026 at 05:48 (UTC+02:00) · discovery channel: web

2. Statnett pauses new grid capacity for data centres in Norway

   Publisher: www.enerdata.net · published 10 November 2026 at 14:00 (UTC+02:00) · retrieved 10 November 2026 at 05:48 (UTC+02:00) · discovery channel: web

3. Norway Halts New Data Center Power Reservations

   Publisher: www.briefs.co · published 9 November 2026 · retrieved 10 November 2026 at 05:48 (UTC+02:00) · discovery channel: web

4. Norway Grid Operator Statnett Halts New Data Center Capacity Reservations

   Publisher: baxtel.com · published 9 November 2026 at 02:00 (UTC+02:00) · retrieved 10 November 2026 at 05:48 (UTC+02:00) · discovery channel: web

5. Norway reins in data centre industry amid AI boom - The Economic Times

   Publisher: m.economictimes.com · published 9 November 2026 at 16:01 (UTC+02:00) · retrieved 10 November 2026 at 05:48 (UTC+02:00) · discovery channel: web

6. Noruega suspende novas conexões de data centers à rede elétrica até 2027 - MegaWhat

   Publisher: megawhat.uol.com.br · published 9 November 2026 at 16:10 (UTC+02:00) · retrieved 10 November 2026 at 05:48 (UTC+02:00) · discovery channel: web

7. Norway Halts Power Connections for New Data Centers

   Publisher: www.bloomberg.com · source publication time unknown · retrieved 10 November 2026 at 05:48 (UTC+02:00) · discovery channel: web

The llama.cpp project announced in GitHub release notes on 10 November that its software can now feed streamed audio to Gemma 4. llama.cpp is a program for running language models on your own computer. The change is in pull request #31965 and is included in build b12237.

With streaming, the model receives audio as a continuous stream rather than a finished file. The change affects the mtmd library, which handles input that accompanies text in llama.cpp, such as audio. The audio stream advances in 480-millisecond steps, the release notes said. They do not mention latency or explain how to enable the audio stream in the server.

The same change includes a fix to the llama.cpp server. The server’s realtime slot forced the model to return embeddings, the model’s internal state, for every row of an input batch. But only a model that generates audio reserves memory for these outputs. As a result, Gemma 4, a chat model using an audio stream, crashed on its very first batch of audio.

The server now turns embeddings on only for models that generate audio. Only the audio generation pipeline reads the embeddings, the release notes said. The notes indicate that Gemma 4 is not an audio-generating model, so the change concerns audio input.

The release comes with prebuilt packages for macOS, iOS, Linux, Android and Windows, and for compute backends such as CUDA, Vulkan, ROCm, SYCL and OpenVINO. The macOS KleidiAI build and the openEuler packages are missing from this release. Users can download the packages from the project’s GitHub page.

What this is based on

1. b12237

   ggml-org · Publisher: github.com · published 10 November 2026 at 01:43 (UTC+02:00) · retrieved 10 November 2026 at 04:18 (UTC+02:00) · discovery channel: GitHub