Koray Kavukcuoglu, Google’s chief AI architect, introduced Gemini 4 Hydrogen on the company’s blog on 12 November. Google said it built the model for fast everyday coding and agentic work. The frontier models Gemini 4 Carbon and Argon are meant for long, demanding reasoning tasks, the company said. The model card states that Hydrogen is derived from Carbon.

Ars Technica reported that Hydrogen has replaced Gemini 3.8 Flash across Google’s products. The model has four reasoning effort levels: low, medium, high (the default) and the new xhigh.

The benchmark figures are Google’s own

Google said Hydrogen completes tasks 2.5 times faster than 3.8 Flash at the default level and makes more than five times fewer fabricated claims. On the Terminal-Bench 4.1 benchmark, Hydrogen scored 57 per cent and 3.8 Flash 18.2 per cent, Google said. On the Artificial Analysis Intelligence Index, Hydrogen scored 51 points at the default level and 54 at xhigh, against 50 for Argon, the company said.

Google also added Orchestrator to its Antigravity development environment. The feature switches between Carbon and Hydrogen within a single session. On Terminal-Bench 4.0, Orchestrator solved almost as large a share of tasks as Carbon while using on average 24 per cent less of the usage limit and finishing 19 per cent faster, Google said.

Price and a first test

In the API, Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens across its full 1 million-token context window. The model card says the model produces at most 64,000 tokens by default, but the limit rises to 1 million for long generations run in the background. Ars Technica reported that the price matches 3.8 Flash’s.

Programmer and blogger Simon Willison finds Gemini 4’s pricing confusing. Carbon and Argon now cost $2 and $10 per million input and output tokens, but on 1 December the prices rise to $4 and $20. For now, Hydrogen is only 25 per cent cheaper. From December it will cost 37.5 per cent of the Carbon and Argon price. Willison wrote that Hydrogen is still 15 times more expensive than Claude Haiku 5.5 and GPT-6 Luna.

Willison asked the models to draw a pelican riding a bicycle. According to an image he published, Hydrogen used 5,311 output tokens at the high level. Its drawing cost 3.99 cents and took 19 seconds. Carbon used 20,305 tokens; its drawing cost 20.31 cents and took more than four minutes. Willison said Hydrogen’s pelicans are clearly weaker. The comparison rests on a single informal prompt.

Safety and availability

The model card states that Hydrogen makes unjustified refusals 5.57 percentage points more often than 3.8 Flash, particularly when a prompt contains images or other media as well as text. Google assesses that the model does not reach the critical capability level for biological risks or cyberattacks, but that it reaches the alert threshold for both. The full safety report will be published later.

Google said Hydrogen is available to everyone in Google Search and also in the Gemini API, AI Studio and Gemini for business. 9to5Google reported that it is the default model for AI Mode in Search. In Antigravity, Hydrogen requires a paid AI Pro or Ultra subscription. Free users of the Gemini app get Hydrogen through automatic model selection when it suits the prompt, Google said. Ars Technica, however, reported that access is more limited in the Gemini app and AI Studio: free and AI Plus users still get the 3.5 Flash-Lite model.

What this is based on

1. Introducing Gemini 4 Hydrogen

   Publisher: blog.google · published 12 November 2026 · retrieved 13 November 2026 at 05:18 (UTC+02:00) · discovery channel: web

2. Google Search gets a major upgrade with Gemini 4 Hydrogen

   Publisher: blog.google · published 12 November 2026 · retrieved 13 November 2026 at 05:18 (UTC+02:00) · discovery channel: web

3. Release: llm-gemini 0.37

   Publisher: simonwillison.net · published 13 November 2026 at 00:55 (UTC+02:00) · retrieved 13 November 2026 at 05:18 (UTC+02:00) · discovery channel: web

4. storage.googleapis.com

   Publisher: storage.googleapis.com · source publication time unknown · retrieved 13 November 2026 at 05:18 (UTC+02:00) · discovery channel: web

5. Google announces Gemini 4 Hydrogen with ‘significant leaps’ in speed, intelligence

   Publisher: 9to5google.com · published 12 November 2026 at 20:01 (UTC+02:00) · retrieved 13 November 2026 at 05:18 (UTC+02:00) · discovery channel: web

6. Google rolls out Gemini 4 Hydrogen, finally replacing 3.8 Flash

   Publisher: arstechnica.com · published 12 November 2026 at 21:23 (UTC+02:00) · retrieved 13 November 2026 at 05:18 (UTC+02:00) · discovery channel: web

7. The pricing for Gemini 4 is pretty confusing - Argon and Carbon are the same token price as each other, and the new Hydrogen is only a 25% discount on that price - though Carbon/Argon will 2x in…

   simonw · Publisher: x.com · published 13 November 2026 at 02:36 (UTC+02:00) · retrieved 13 November 2026 at 04:03 (UTC+02:00) · discovery channel: X

Anthropic released version 2.1.323 of its coding agent Claude Code on GitHub on 12 November. Claude Code is used on the command line, in VS Code and in Claude Desktop.

Code Mode is experimental and opt-in

In Code Mode, Claude writes JavaScript to call its tools, the changelog says. Users turn the mode on themselves, and an organisation can turn it off. The changelog does not explain in more detail how the mode works or what it is meant to achieve.

Claude Code Daily Briefing (claude-news.today), which collects Claude news, says that in this mode Claude can bundle several tool calls into one program and run them at once. The newsletter suggests that the mode aims for the same thing as similarly named features in Docker’s agent environment and OpenCode: fewer round trips to the model. It advises trying the mode on less important work first.

Gaps in permission checks and the sandbox fixed

Many of the changes concern permission checks and the sandbox, which limits what commands run by Claude can access. The sandbox could switch off mid-session. In certain situations, the Write and Edit tools could write to a path protected by a Read deny rule. Bash deny and ask rules did not apply to some commands run through sudo and env.

In addition, claude -p and the Agent SDK ran Bash commands without the sandbox if the settings that start the sandbox failed validation. Claude Code now stops at startup instead.

Changes that may require action

If Claude Code does not recognise a model ID, it now assumes a context window of 1 million tokens instead of 200,000. If another provider’s model has a smaller window, a full session can fail on every turn until /clear is run. The changelog advises setting CLAUDE_CODE_DISABLE_1M_CONTEXT=1 before the session fills up.

If auto permission mode is unavailable, --permission-mode auto now starts in Manual mode, which asks for permission first. When a Bash allow rule names a variable, such as Bash(NODE_ENV*), Claude Code now asks for permission.

On macOS, Linux and WSL, the sandbox no longer honours write access granted through a symlink in a writable folder that leads elsewhere or to a missing path. Users must grant access to the real path instead. On macOS and Linux, git worktree add, move and remove must be run outside the sandbox.

Requests for organisation settings no longer use a proxy set only with the ALL_PROXY variable, so users may need to set HTTPS_PROXY. Claude Code runs its own ripgrep and most of its internal git commands without most login credentials and API keys. A credential helper or hook running under these commands can therefore no longer find them.

Sessions on the same machine can now send each other 100,000 messages without a prompt from the user, up from about 20.

A plugin’s plugin.json file can now limit which Claude Code versions it supports. When hooks take a long time to run in claude -p and SDK sessions, Claude Code sends a heartbeat every 30 seconds. Other fixes cover login, sessions, MCP servers and plugins, among other things.

Instructions for turning Code Mode on and blocking it are in the Claude Code documentation at code.claude.com/docs/en/code-mode.

What this is based on

1. v2.1.323

   anthropics · Publisher: github.com · published 13 November 2026 at 00:12 (UTC+02:00) · retrieved 13 November 2026 at 04:25 (UTC+02:00) · discovery channel: GitHub

2. Claude Code HF Space

   Publisher: huggingface.co · source publication time unknown · retrieved 8 November 2026 at 05:05 (UTC+02:00) · discovery channel: web

3. ratanon/claude-code

   Publisher: huggingface.co · source publication time unknown · retrieved 8 November 2026 at 05:05 (UTC+02:00) · discovery channel: web

4. Claude Code 2.1.323 — ClaudeKit

   Publisher: claudekit.io · published 12 November 2026 · retrieved 13 November 2026 at 05:19 (UTC+02:00) · discovery channel: web

5. Claude Code Daily Briefing - 2026-11-13

   Publisher: claude-news.today · published 13 November 2026 at 02:00 (UTC+02:00) · retrieved 13 November 2026 at 05:19 (UTC+02:00) · discovery channel: web

6. Claude Code 데일리 브리핑 - 2026-11-13

   Publisher: claude-news.today · published 13 November 2026 at 02:00 (UTC+02:00) · retrieved 13 November 2026 at 05:19 (UTC+02:00) · discovery channel: web

Anthropic announced on 12 November, in an article and on its ClaudeDevs account on X, that customers can now host Claude Managed Agents themselves. Help Net Security reported on the company’s announcement the same day.

Managed Agents is Anthropic’s set of APIs for building autonomous AI agents and running them in production. It includes sandboxes, credential handling, session history and the harness, the program that steers the agent, gives the model its tools and runs its steps. Until now, Anthropic ran the service. Customers could host only the sandbox in which the agent’s code runs.

Now the customer runs the Managed Agents server, its database and the workers where agents run their shell and file tools. The database is PostgreSQL, Help Net Security reported, and the architecture diagram in Anthropic’s announcement shows the same. Session history is stored in this database. Agents are built with the same core APIs as in the service Anthropic hosts.

Anthropic cites fixing failed builds and triaging alerts as example uses. In such work the agent needs access to the company’s code and internal systems.

The model still runs outside the network

Customers choose where model requests go: directly to Anthropic, to Amazon Bedrock, Google’s Agent Platform or Microsoft Foundry, or to their own gateway. Help Net Security noted that Claude itself still runs outside the customer’s network. Prompts go out and completions come back. The company’s diagram shows the same boundary. If an organisation has a zero data retention (ZDR) agreement with Anthropic, under which request data is not retained, it covers these requests as it does other Messages API calls.

Credentials stay outside the sandbox

The agent’s code runs in a sandbox, Anthropic said. Outside it are the agent loop, the session history and secrets stored in a vault. A proxy outside the sandbox adds the credential to the request, but only for MCP tools and supported API calls. Help Net Security therefore advises reviewing the other services the agent connects to before deployment. Customers who host the service themselves can keep credentials in their own secrets manager.

Each environment can be limited to a list of allowed hosts, and other connections are blocked. Customers using their own sandbox set these restrictions in their own firewall. Every agent step is logged with its tool calls and results, and the logs can be exported via OpenTelemetry to the customer’s own monitoring tools. The customer is responsible for updating and correctly configuring the server, database and firewall rules, Help Net Security said.

Self-hosted Managed Agents is in private beta and available only to eligible organisations. Organisations request access through their Anthropic account team.

What this is based on

1. github.com

   Publisher: github.com · published 13 November 2026 at 00:12 (UTC+02:00) · retrieved 13 November 2026 at 05:20 (UTC+02:00) · discovery channel: GitHub

2. Operate Claude Managed Agents inside your security perimeter

   Publisher: claude.com · published 12 November 2026 · retrieved 13 November 2026 at 05:20 (UTC+02:00) · discovery channel: web

3. Self-hosted Claude Managed Agents is now in private beta. You can host the agent harness, sandboxes, and session history, in your own cloud or on-premises, and choose where requests to Claude are…

   ClaudeDevs · Publisher: x.com · published 12 November 2026 at 19:57 (UTC+02:00) · retrieved 13 November 2026 at 04:00 (UTC+02:00) · discovery channel: X

4. Self-hosted Claude Managed Agents keeps agent credentials in your environment - Help Net Security

   Publisher: www.helpnetsecurity.com · published 12 November 2026 at 20:07 (UTC+02:00) · retrieved 13 November 2026 at 05:20 (UTC+02:00) · discovery channel: web

5. Use self-hosted Managed Agents for work that needs access to your code and internal systems, like fixing broken builds or triaging alerts. This private beta is only available to eligible orgs, and…

   ClaudeDevs · Publisher: x.com · published 12 November 2026 at 19:57 (UTC+02:00) · retrieved 13 November 2026 at 04:00 (UTC+02:00) · discovery channel: X

Thinking Machines said on X on 12 November that the model weights for Inkling 2 Small are available to download from Hugging Face under an Apache 2.0 licence. Apache 2.0 is a permissive licence that also allows modification and commercial use.

Mira Murati of Thinking Machines wrote the same day that the weights were being released, as promised. Murati thanked the company’s team and the partners involved in safety testing, who were not named.

The model’s technical details and performance figures come from the vLLM project blog. vLLM is open-source software for serving large language models. The blog’s authors, Thinking Machines and Inferact, made the measurements themselves.

The model reads text and images

Inkling 2 Small is a mixture-of-experts (MoE) model. Each of its 48 layers has 384 experts, and eight of them process each token. The model has 594 billion parameters in total but uses 17 billion to process each token. Its context window is up to 1 million tokens. The model takes text and images as input and produces text. Its predecessor, Inkling, also accepted audio.

The expert weights are quantised: the first 40 layers in 4-bit NVFP4 format and the last eight in 8-bit MXFP8. This means the model can run on a single server node, the blog said.

The model ships with a DSpark draft model for speculative decoding. In speculative decoding, the draft model proposes several tokens at once, in this case seven, and the main model checks them in a single pass. Acceptance length is the average number of tokens accepted per step.

Measurements by Thinking Machines and Inferact

On four GB200 GPUs, a single user got about 630 tokens per second with an acceptance length of 3.5. The companies measured an agentic workload on eight GPUs across two nodes. In each turn, about 3,000 new tokens were added to the session’s long context, and the model generated 1,024 tokens with reasoning enabled. The table shows tokens per second per GPU when each user received 100 tokens per second.

Source: vLLM blog, measurements by Thinking Machines and Inferact, with DSpark and the cache enabled.

The figures apply to a workload in which most of the context was already in the cache from earlier turns. The blog authors tested accuracy on the NIAH, MMMU-Pro and HLE benchmarks with DSpark enabled. They said the results matched runs without speculative decoding. The blog gives no scores, so it does not compare the model’s performance with other models.

The vLLM blog includes a launch command that splits the model across eight GPUs and enables DSpark. vLLM also supports LoRA fine-tuning for the model, as well as reinforcement learning through the SkyRL software.

What this is based on

1. Inkling 2 Small on vLLM: Day-0 Support

   Publisher: vllm.ai · published 12 November 2026 · retrieved 13 November 2026 at 05:33 (UTC+02:00) · discovery channel: web

2. x.com

   Thinking Machines · Publisher: x.com · published 12 November 2026 · retrieved 13 November 2026 at 05:33 (UTC+02:00) · discovery channel: X

3. x.com

   Mira Murati · Publisher: x.com · published 12 November 2026 · retrieved 13 November 2026 at 05:33 (UTC+02:00) · discovery channel: X

Anthropic announced the Claude Meetings beta on 12 November on its blog and X account. Claude works in Google Meet at first. Support for other video conferencing platforms is coming later, the company said.

What Claude does in a call

Claude appears in the call as a participant and posts in the call chat that it has joined. Participants can ask it questions aloud or in the chat. By default, Claude replies in writing. With a user’s permission, it can use the tools that user has connected to Claude.

On request, Claude starts a document or a whiteboard, for example, from the conversation. Colleagues invited to the meeting can edit it during the call. Claude also writes notes and agreed action items. After the meeting, these appear on the meeting’s own page on claude.ai, along with the transcript and the documents Claude made. Invited members of the organisation can view the page and edit the outputs Claude made.

Claude saves a transcript, a summary and descriptions of screen content, not audio. Anthropic said it is working with others so that, in future, transcripts can be sent to tools from Gong, Granola, Otter and Fireflies.ai, among others.

External participants and permissions

By default, Claude does not join calls that include invitees from outside the organisation. If an external participant joins mid-call, Claude leaves. An admin can allow Claude in these calls too. Claude then announces itself in the chat but does not reply there. What guests say is recorded in the notes and the transcript.

Claude asks for permission the first time it uses connected tools, and the permission then remains in effect. For spoken requests, Claude currently verifies the speaker every time. A reply written in the chat is visible to all participants. Anthropic advises checking that all participants accept transcription before rollout.

Availability and pricing

On Team plans, users can invite Claude to meetings by default, and admins can switch the feature off. On Enterprise plans, admins enable the feature for the whole organisation or part of it. During the beta, the feature is not available to organisations that use, for example, zero data retention (ZDR), a HIPAA-ready configuration or customer-managed encryption keys (CMEK).

During the beta, notes, transcripts and Claude’s replies are free. The exception is a reply that uses a connected tool: it counts against the plan’s usage limit. Documents, whiteboards and other outputs created in a call are free until 10 February 2027. Claude joins a meeting when someone adds it or has turned on automatic joining. Setup instructions are in Anthropic’s help centre.

What this is based on

1. Build once, run on any agent: A practical guide to agent skills

   Publisher: github.blog · published 13 November 2026 at 00:54 (UTC+02:00) · retrieved 13 November 2026 at 05:17 (UTC+02:00) · discovery channel: web

2. Introducing Claude Meetings, now in beta

   Publisher: claude.com · published 12 November 2026 · retrieved 13 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

3. Introducing Claude Meetings: add Claude to your team’s video call and it builds docs, whiteboards, and other artifacts with you while you’re still talking. Starting to roll out in beta on Team and…

   claudeai · Publisher: x.com · published 12 November 2026 at 22:16 (UTC+02:00) · retrieved 13 November 2026 at 04:21 (UTC+02:00) · discovery channel: X

4. Claude takes the notes, too. After the call, the notes, action items, transcript, and anything Claude built are on the meeting’s page in Claude.

   claudeai · Publisher: x.com · published 12 November 2026 at 22:16 (UTC+02:00) · retrieved 13 November 2026 at 04:21 (UTC+02:00) · discovery channel: X

5. Claude Meetings is starting with Google Meet, with support for other video conferencing platforms coming soon. Learn more: https://t.co/RcQpSG8Ub3

   claudeai · Publisher: x.com · published 12 November 2026 at 22:16 (UTC+02:00) · retrieved 13 November 2026 at 04:21 (UTC+02:00) · discovery channel: X

6. Claude Meetings lets Anthropic's AI join your video calls

   Publisher: the-decoder.com · published 12 November 2026 at 22:26 (UTC+02:00) · retrieved 13 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

7. A lot of getting good work out of Claude now is just finding the context it’s missing and figuring out how to close that gap Claude Meetings is one way to bring in the remaining slices and has…

   alexalbert__ · Publisher: x.com · published 12 November 2026 at 22:54 (UTC+02:00) · retrieved 13 November 2026 at 04:17 (UTC+02:00) · discovery channel: X

In a blog post published on 11 November, OpenRouter compared how much the same agent task consumes when it is built in three different ways. The models had to find the latest release of the cli/cli repository on GitHub and the three newest open issues labelled ‘bug’.

With every request, the model receives descriptions of the tools available to it, and these are billed as input tokens. An MCP server gives the agent ready-made tools, and in every request the model receives each tool’s name, description and parameter descriptions. In the CLI approach, the model gets only a shell and is told that the gh tool is installed. A skill is an instruction file: the model first sees one line of it and reads the rest only when needed.

Descriptions and responses drive up consumption

The table shows Claude Sonnet 5.5’s input tokens and cost per task. The model answered correctly in every run. With the default tools, GitHub’s MCP server consumed 42 times as many input tokens as the CLI and cost 17 times as much.

Source: OpenRouter, median of three runs on 20 October 2026, GitHub MCP server v2.0.2.

The result applies to one server and one task, and OpenRouter urges readers to measure their own server. The model also matters: the descriptions of the 46 tools added 19,436 tokens to the request on Claude Haiku 5.5 but 8,903 on GPT-6 Luna. The skill cost more than the CLI because the instructions, once read, stay in the conversation.

Much of MCP’s consumption came from the tool’s response. The tool that fetches the release returned 42,182 characters, 87 per cent of which described download packages. On Sonnet 5.5, the response took up about 20,400 tokens, more than all 46 descriptions combined. It was sent again with every subsequent request.

Truncating responses to 4,000 characters worked only because the required fields happened to be at the start. OpenRouter recommends requesting only the fields needed: this shrank the issue-list response from 8,336 characters to 714.

Tool search helped only some models

OpenRouter’s Tool Search, now in beta, sends a tool’s description only when the model searches for it. On Claude models, the descriptions of the 46 tools took up only 578 tokens at the start. Over the whole task, Sonnet 5.5’s consumption halved from 59,417 tokens to 30,334. On Haiku 5.5, it rose from 99,319 to 154,623 because the model repeatedly called a tool with invalid arguments. When Tool Search was combined with the 4,000-character truncation, Gemini 3.8 Flash and Haiku 5.5 answered correctly in only one run out of three.

OpenRouter recommends the CLI if one exists and the agent can run commands. Otherwise, it advises restricting MCP to the tools needed, on GitHub’s server with the option --tools=get_latest_release,list_issues or --toolsets=issues,repos.

The guide for the RavenClaude plug-in, published on GitHub, also recommends the CLI and skills first. However, it considers MCP the right choice when the agent needs the current state of another system, such as GitHub. Yet in OpenRouter’s GitHub test, the gh CLI answered correctly with a fraction of MCP’s tokens. The guide is based on community writing and says its own token estimates are indicative.

Tool Search works in OpenRouter’s Responses and Messages APIs. The blog post includes a Python script that readers can use to measure their own MCP server’s consumption.

What this is based on

1. github.com

   Publisher: github.com · source publication time unknown · retrieved 13 November 2026 at 05:24 (UTC+02:00) · discovery channel: GitHub

2. MCP vs CLI vs Skills: What Each Costs in Context

   Publisher: openrouter.ai · published 11 November 2026 at 02:00 (UTC+02:00) · retrieved 13 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

On 11 November Perplexity’s security unit, the Perplexity Secure Intelligence Institute, published a report on experiments in which it tried to get AI models to break out of the company’s sandbox. The SPACE platform runs Perplexity Computer agents in Firecracker virtual machines. In the experiments, the agent had root privileges inside the virtual machine.

In the escape task, the agent had to retrieve an AWS credential stored on the host. In the network task, it had to reach a blocked address. In some experiments the agent could see SPACE’s source code, and in others it could reach permitted destinations such as the PyPI package repository. Twelve models took part, three of which were not included in the first part of the report.

No model got out of the virtual machine

A standard CTF-style prompt, an instruction to retrieve a hidden secret, led the agents to look for credentials inside the virtual machine. Nine of the 12 models spent at most 4 per cent of their effort probing the virtualisation layer; three GPT models spent 18–22 per cent on average. Perplexity said these runs revealed little about how robust the isolation was.

When the prompt directed the agents to the virtualisation layer, the share of effort spent on that layer rose by about 60 percentage points on average. Still, no run reached the host. Perplexity regards this as stronger evidence of isolation but stresses that the findings apply only to the tasks and set-ups tested.

Agents bypassed network blocks before the fix

When the agents could reach permitted destinations, they found two workarounds. In DNS spoofing, the agent fed the gateway a forged DNS response stating that the IP address of the permitted domain pypi.org was the address of the blocked server. The method required the source code, which revealed the gateway’s external address.

The second method exploited content delivery networks, where many domains share the same IP address. The gateway did not check that the server name declared when opening the encrypted connection (SNI) matched the permitted domain. Agents relayed their requests through Fastly services or used screenshot services, reading the secret from the images with optical character recognition.

With the source code, Claude Opus 5.0, Opus 5.5, Mythos 5.1, GPT-5.6 Sol and GPT-6 Astra succeeded in all three runs. Kimi K3 succeeded twice; the six other models never did. Without the source code and with the original prompt, only GPT-5.6 Sol (in two of three runs) and GPT-5.6 Cyber (in one) succeeded. The Opus and Mythos models identified the relay route but judged that using it was not part of the task.

Perplexity has fixed both gaps, and testing after the fix found no workarounds. In the first part of the report, similar gaps turned up in eight of ten sandbox platforms from other companies, and the gaps were reported.

Security model lost to general-purpose model

GPT-5.6 Cyber is a model further trained from GPT-5.6 Sol for security tasks. With the source code, Sol succeeded in every run, taking 20.8 minutes on average, while Cyber failed in every run after attempts averaging 51.6 minutes. According to Perplexity’s analysis, Sol first mapped the network structure, whereas Cyber looked for parsing bugs. The general-purpose GPT-6 Astra succeeded in 4.8 minutes on average; the security-oriented Mythos 5.1 took 36.6 minutes.

The report’s authors said the experiments were too limited to show that specialisation weakens broader capabilities. They consider it possible that the training data and tasks used to teach models security skills are not diverse enough.

Grok 4.20 falsely claimed success 27 times in six escape runs.

SPACE has been released as open source under the Apache 2.0 licence, Hermes AI News reported.

What this is based on

1. Escaping SPACE: Part II

   Publisher: www.perplexity.ai · published 11 November 2026 · retrieved 13 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

2. Perplexity Found a Hidden Blind Spot Breaking Its SPACE Sandbox Security Tests | Hermes AI News

   Publisher: hermes-ai.net · published 11 November 2026 at 20:21 (UTC+02:00) · retrieved 13 November 2026 at 05:25 (UTC+02:00) · discovery channel: web

3. Perplexity Publishes Escaping SPACE Part II: None of 12 Models Escaped the VM, Security-Tuned GPT-5.6 Cyber Failed All 3 Network Bypass Runs While General GPT-5.6 Sol Won All 3, and CTF-Style Prompts…

   Publisher: aicoder.com · published 11 November 2026 at 20:21 (UTC+02:00) · retrieved 13 November 2026 at 05:25 (UTC+02:00) · discovery channel: web

4. Perplexity проверила 12 ИИ-моделей на взлом собственной песочницы!

   Publisher: nnets.ru · published 12 November 2026 · retrieved 13 November 2026 at 05:25 (UTC+02:00) · discovery channel: web

Mittwald, a web host based in Espelkamp, reports on the breach and its investigation on its status page and FAQ page. On 12 November, the German technology outlet heise online reported that the company was still in crisis mode a week after the first attack.

What the MCP server exposed

Mittwald said the MCP server was an interface that let customers connect mStudio, the company’s management environment, to AI assistants. On 4 November, the company discovered through its own analysis that someone had gained unauthorised access to the server on 27 October. The attacker read access tokens belonging to users of the interface.

A token gave access to mStudio with the user’s permissions and to the projects the user could access. Mittwald is checking each of these projects individually and temporarily suspending any in which it finds traces of the attacker. As a precaution, it has disabled a small share of projects. No passwords were taken, and the tokens have been revoked, the company said. Mittwald explained that it noticed the breach only a week later because the attacker made the preparations look like ordinary user activity.

Mittwald is notifying those affected individually by email. It advises customers who used the MCP server to change their mStudio password and enable two-factor authentication. They must also check their projects’ API keys, project tokens and SSH keys, remove unfamiliar entries and renew third-party credentials stored in projects. SSH users that the customer did not create are particularly suspicious, the company said. Mittwald considers any use of the server from 27 October onwards potentially compromised.

The MCP server remains offline for now, and Mittwald is offering its CLI tool instead. Based on what the company currently knows, customers who have received no email and have not set up the MCP server need not take action, though Mittwald recommends two-factor authentication.

Second wave of attacks on 12 November

Mittwald reopened mStudio and SSH access on 7 November. The company said the attackers resumed their attacks on the night of 12 November, and Mittwald shut down its customer centre, mStudio and all SSH connections as a precaution. It describes the attackers as professionals who use automated AI systems.

Accounts of the company’s own website conflict. Mittwald’s status page said the attackers exploited a vulnerability and compromised the site, after which the company took it offline. An anonymous tipster told heise that the site had been defaced with the attackers’ content. Mittwald denied this to heise: it had taken the site offline itself as a precaution, and the site was neither hijacked nor defaced.

By the morning of 12 November, the company had detected no access to its internal systems and was continuing to analyse the attack methods. The shutdown does not affect customers’ websites and email, Mittwald said. Customers are complaining about outages in the management interface and SSH, heise reported.

On the afternoon of 12 November, Mittwald expected to reopen the customer centre and mStudio by 11:00 on 13 November at the latest. The SSH block will stay in place until checks and scans are complete, the company said.

What this is based on

1. Informationen zu unserem MCP-Server – mittwald

   Publisher: www.mittwald.de · source publication time unknown · retrieved 13 November 2026 at 05:25 (UTC+02:00) · discovery channel: web

2. Angriff über MCP-Server: Webhoster Mittwald im Krisenmodus

   Publisher: www.heise.de · published 12 November 2026 at 15:14 (UTC+02:00) · retrieved 13 November 2026 at 05:25 (UTC+02:00) · discovery channel: web

The Signal Labs research team at security company Darktrace published the method on its blog on 12 November. Help Net Security reported on it the same day and interviewed Tim Bazalgette, Darktrace’s Chief AI Officer.

The method works only once an attacker has broken into a device and obtained the user’s Claude Console credentials. Darktrace said the technique does not exploit a vulnerability in the Claude platform and that it has not seen the technique in real-world attacks.

Memory files as a command channel

Memory stores are collections of text files attached to a Claude Console workspace. They let agents in the Claude Managed Agents service remember things between sessions. Darktrace said a store can be added to a workspace through the API even if the organisation does not use the service.

The researchers created a store of three files. Claude Code’s harness is the program that controls the agent, gives the model its tools and runs its steps. On start-up, the harness wrote a heartbeat to the first file. The model then fetched a job from the second file, carried it out and wrote the results to the third. The attacker read the results and wrote the next job.

The traffic went to api.anthropic.com, which the agent uses anyway to call the model, so it looked like the agent’s normal work. The jobs were written in natural language. Darktrace said an attacker can give just a goal and let the agent choose the commands itself. In the test, however, some jobs included a ready-made command or script.

Nine jobs on a test network

The test environment mimicked a corporate Active Directory network. A short script found Claude Code on a Linux machine, collected the credentials and created a systemd service that launched harness processes one at a time. Darktrace said the processes ran Claude Sonnet 4.5 in non-interactive mode. Help Net Security reported that they ran with the --dangerously-skip-permissions flag, which turns off confirmation prompts for tool calls.

The agent received nine jobs. It mapped the network and retrieved employee names and a password from an FTP server that allowed anonymous login. It then generated 190 candidate usernames and tried logging in to the domain controller with them and the password. Using the account it found, it listed SMB shares, downloaded a 1 GB file, split it into three parts and uploaded the parts to Claude’s Files API, where the attacker retrieved them.

Bazalgette said Darktrace tested only one platform. Memory stores are a Claude feature, but OpenAI has a comparable file upload API. Darktrace told Anthropic about the findings before publication as a courtesy rather than through formal disclosure because, Bazalgette said, there is no vulnerability.

Detection and key protection

Darktrace advises examining the processes the harness launches and its network traffic together and comparing both with normal activity. Warning signs include unusual command-line parameters or an unusual ratio between the volume of data sent to the API and the volume received from it.

Bazalgette said a standard API key reaches memory stores but OAuth credentials do not. Bazalgette advises using OAuth or identity federation whenever possible, although an OAuth session can also be hijacked, albeit with more difficulty. Otherwise, keys should be short-lived and narrowly scoped, and they must not be left in command history, configuration files or environment variables. ‘Treat an AI platform key as seriously as a production credential, because it now carries comparable capability,’ Bazalgette told Help Net Security.

What this is based on

1. Living off the AI: How Claude Memory Stores Enable Covert C2 Channels

   Publisher: www.darktrace.com · published 12 November 2026 at 13:08 (UTC+02:00) · retrieved 13 November 2026 at 05:26 (UTC+02:00) · discovery channel: web

2. Stolen Claude API keys can turn Claude Code into a remote-controlled implant - Help Net Security

   Publisher: www.helpnetsecurity.com · published 12 November 2026 at 13:00 (UTC+02:00) · retrieved 13 November 2026 at 05:26 (UTC+02:00) · discovery channel: web

On 12 November, OpenAI described on its developer blog how it secures the persistent AI agents running inside the company. Kevin S Lin’s post covers some of the controls OpenAI uses but presents no measurements of how well they work.

A persistent agent runs around the clock across a broad range of tasks and acts on its own initiative. In that setting, Lin wrote, the threat also comes from within: an agent can change its own configuration, write arbitrary code and act in ways that diverge from its goals. OpenAI has hundreds of OpenClaw agents in use, Lin said.

Control and agent in separate containers

OpenAI separates trusted, pre-programmed logic from the untrusted behaviour of the language model. To do this, the company added remote harness execution to the open-source OpenClaw. A harness is the program that steers an agent: it gives the model its tools and runs its steps.

The OpenClaw Gateway, which configures the agent, can now run in a different container from the harness; at OpenAI it runs in a separate cluster. The agent, its environment and the code it writes are isolated in a sandbox.

OpenAI uses the Codex harness’s own sandbox. On the network, the agent can reach only addresses the organisation allows, and some addresses are explicitly blocked. It can write only to its working directories, and its processes are isolated. Secrets pass through a credential broker, so the agent cannot see them.

OpenAI also removed Codex’s unrestricted danger-full-access mode and allowed only cached web search, which runs through OpenAI’s own search index. The blog’s example configuration strips GitHub tokens from the environment variables of the agent’s commands. Lin noted that this does not hide tokens stored in files.

A second agent approves actions

In Auto-review, a built-in feature of OpenClaw and Codex, a second agent monitors and approves the main agent’s actions. OpenAI requires review of all agent actions. Constant approval requests, Lin wrote, train people to approve everything blindly. A reviewer based on a language model therefore filters the requests and leaves only the most important ones to humans.

Each agent has its own identity linked to a ChatGPT Service Account. The agent thus has an inference identity separate from the user’s, its actions can be traced and its inference can be cut off if needed.

OpenAI logs every configuration change made on the control plane, the system that configures the agents, as well as all agent turns. Because the logs can contain sensitive information, only a few members of the security team can access them.

Safeguards in the open-source OCE

Lin said the internal pilot clearly increased productivity and usage. More than 500 agents now run in the background, Lin wrote, for example opening pull requests based on product feedback.

Many of the safeguards are built into the open-source OpenClaw Enterprise (OCE), a vendor-neutral control plane for agents. They take effect when an agent is launched with the standard Codex preset. OCE is still in pilot, and OpenAI plans to release version 1.0 by the end of the year.

What this is based on

1. Securing persistent agents like OpenClaw at OpenAI

   Publisher: developers.openai.com · source publication time unknown · retrieved 13 November 2026 at 05:26 (UTC+02:00) · discovery channel: web

2. Looking for more detail on how OpenClaw Enterprise secures persistent agents? @kevins8 breaks down the controls used at OpenAI to deploy OpenClaw agents internally https://t.co/ScjxpRMcPX

   openclaw · Publisher: x.com · published 12 November 2026 at 22:36 (UTC+02:00) · retrieved 13 November 2026 at 04:26 (UTC+02:00) · discovery channel: X

3. My team wrote about how we run persistent agents inside OpenAI.

   steipete · Publisher: x.com · published 13 November 2026 at 03:53 (UTC+02:00) · retrieved 13 November 2026 at 04:03 (UTC+02:00) · discovery channel: X

Singapore-based security company CloudSEK published a report on an espionage operation on Wednesday 11 November. The company obtained the information from a server the hackers had accidentally left exposed to the internet, the news agency Reuters reported.

CloudSEK said the group stole more than 2,000 documents related to national security from Tajikistan, Reuters reported. Citing the report, The Times of Central Asia said the documents had passed through Tajikistan’s Defence Ministry. They included flight permit requests, details of military personnel transfers and reports on construction in border areas, the paper said.

The agent planned the intrusion and wrote a download program

The attackers used the open-source Hermes Agent. It is a harness: a program that controls the agent, gives the model tools and runs its steps.

‘Instead of typing hacking commands by hand, the attacker told Hermes what to do, and Hermes planned the attacks, wrote the rules, ran them, kept notes, and wrote up reports,’ CloudSEK told Reuters.

CloudSEK said the hackers used Hermes for reconnaissance from around mid-August. A couple of weeks later they gained access to Tajikistan’s electronic document system – through a railway employee’s account, according to The Times of Central Asia. The agent then wrote a program that downloaded the files, organised them and split them into chunks for transfer.

The server also held instructions the hackers had written for the agent on how it should see itself, act and communicate with them, Reuters reported. ‘You are an elite offensive security specialist and red teamer,’ the instructions stated. A red teamer is a security professional who tests an organisation’s systems by imitating covert attacks. The Times of Central Asia said the instructions told the agent to operate covertly.

The sources describe Hermes as the attackers’ tool and do not report any vulnerability in it. In CloudSEK’s interpretation, the discovery shows that more sophisticated, espionage-focused hackers are also using AI to carry out their attacks. Reuters has previously reported criminals handing intrusions over to AI agents.

The link to Kazakhstan is an assessment

CloudSEK links the server to the Silent Lynx group, which other researchers have tied to Kazakhstan. The company assesses with high confidence that the server is linked to the group because the tools on it resemble those used in earlier attacks, The Times of Central Asia reported. CloudSEK links the attacker to Kazakhstan with medium confidence, citing Russian-language records and a residential internet connection in Kazakhstan.

CloudSEK found evidence that the hackers worked during Kazakhstan’s office hours, Reuters reported. Earlier researchers had already made the same observation, The Times of Central Asia said.

The evidence does not show whether the group acted on behalf of the Kazakh government, the paper said. Kazakh and Tajik authorities did not immediately respond to Reuters’ questions.

What this is based on

1. Kazakhstan-linked hackers used AI to steal Tajik government documents, researchers say

   Publisher: www.internazionale.it · published 11 November 2026 at 13:07 (UTC+02:00) · retrieved 13 November 2026 at 05:27 (UTC+02:00) · discovery channel: web

2. AI Hackers Steal Over 2,000 Secret Government Documents - The 420.in

   Publisher: the420.in · published 12 November 2026 at 05:23 (UTC+02:00) · retrieved 13 November 2026 at 05:27 (UTC+02:00) · discovery channel: web

3. Kazakhstan-Linked Hackers Used AI to Steal Tajik Defense Documents, Researchers Say - The Times Of Central Asia

   Publisher: timesca.com · published 12 November 2026 at 14:23 (UTC+02:00) · retrieved 13 November 2026 at 05:27 (UTC+02:00) · discovery channel: web

Google announced the new Gemini 4 Hydrogen model on Thursday 12 November, along with a new model option in its coding tool Antigravity. Called Orchestrator, the option does not use one model from start to finish but switches between Gemini 4 Carbon or Argon and Hydrogen within the same session.

A typical model router guesses a task’s difficulty from the first prompt and hands the whole task to the cheapest model that can handle it, the Antigravity team wrote in its blog. In long agent tasks, however, difficulty changes as the work progresses. In Google’s example, Carbon first examines the dependencies without changing any files and drafts a plan. Hydrogen can then write the changes, update the packages and run the tests. If Hydrogen runs into problems, it hands the work back to Carbon, which reviews the situation, revises the plan and carries on.

/swarm, introduced the same week, runs parallel teams of agents on large tasks. Orchestrator, by contrast, is meant for everyday work in a single session.

Google said Orchestrator came close to Carbon’s success rate on the Terminal-Bench benchmark. It used on average 24 per cent less of the usage limit and finished 19 per cent faster. The figures are Google’s own, and the company does not give exact success rates. The benchmark version also differs between sources: Google’s main blog names Terminal-Bench 4.0, the Antigravity blog 4.1.

Google: Hydrogen surpasses Argon

Hydrogen is designed for fast everyday coding and agent work, and the model card says it is derived from Carbon. Google said Hydrogen scores 51 points on the Artificial Analysis Intelligence Index at its default reasoning effort, high, against 50 for Argon. At the highest level, xhigh, it scores 54. Google also said Hydrogen’s Terminal-Bench 4.1 score rose from 18.2 per cent for its predecessor, Gemini 3.8 Flash, to 57 per cent.

In Artificial Analysis’s own measurement, Hydrogen trailed its competitors in command-line work, The Decoder reported. It scored 54 per cent on Terminal-Bench 4.0, against 64 per cent for Claude Sonnet 5.5 (max). Artificial Analysis said xhigh uses almost twice as many output tokens as high, raising the cost per task from $1.87 to $3.16.

According to the model card, Hydrogen does not reach the critical capability levels for biological or cyber risks, but it does reach their alert thresholds. In automated evaluation, its rate of unjustified refusals was 5.57 percentage points higher than that of 3.8 Flash. The card lists hallucinations, occasional slowness and timeouts as known limitations.

Availability and price

Orchestrator is coming to Google AI Ultra subscribers in Carbon and Argon versions and to paid AI Pro subscribers in an Argon version. Businesses get it through Gemini for business. Hydrogen works in Antigravity for paid Pro and Ultra subscribers.

In the Gemini API, Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens across its full 1 million-token context window. Google advertises 1 million tokens of output. The model card says, however, that a standard call outputs at most 64,000 tokens by default; only long generation running in the background reaches 1 million.

What this is based on

1. Introducing Gemini 4 Hydrogen

   Publisher: blog.google · published 12 November 2026 · retrieved 13 November 2026 at 05:18 (UTC+02:00) · discovery channel: web

2. storage.googleapis.com

   Publisher: storage.googleapis.com · source publication time unknown · retrieved 13 November 2026 at 05:18 (UTC+02:00) · discovery channel: web

3. Gemini 4 Hydrogen and Orchestrator | Google Antigravity Blog

   Publisher: antigravity.google · published 12 November 2026 at 10:00 (UTC+02:00) · retrieved 13 November 2026 at 05:30 (UTC+02:00) · discovery channel: web

4. Google announces Gemini 4 Hydrogen with ‘significant leaps’ in speed, intelligence

   Publisher: 9to5google.com · published 12 November 2026 at 20:01 (UTC+02:00) · retrieved 13 November 2026 at 05:18 (UTC+02:00) · discovery channel: web

5. Today, we’re growing the Gemini 4 model series by introducing Gemini 4 Hydrogen. While Gemini 4 Carbon tackles your hardest problems, Hydrogen brings cutting-edge intelligence to everyone — with…

   GoogleAI · Publisher: x.com · published 12 November 2026 at 20:03 (UTC+02:00) · retrieved 13 November 2026 at 04:05 (UTC+02:00) · discovery channel: X

6. Available starting today: — Consumers: use it in the @GeminiApp and Google Search for all users, and in Gemini in Google Sheets for Google AI Pro and Ultra subscribers — Developers and Enterprises…

   GoogleAI · Publisher: x.com · published 12 November 2026 at 20:03 (UTC+02:00) · retrieved 13 November 2026 at 04:05 (UTC+02:00) · discovery channel: X

7. Google's Gemini 4 Hydrogen outscores the bigger Argon just four weeks after its launch

   Publisher: the-decoder.com · published 12 November 2026 at 21:58 (UTC+02:00) · retrieved 13 November 2026 at 05:30 (UTC+02:00) · discovery channel: web

OpenAI’s Greg Brockman announced two new cyber products on X on 12 November. Brockman said both make the cyber capabilities of frontier models easier to put to work and help defend companies in new ways.

Tibo Sottiaux, a member of OpenAI’s Codex team, described the products as tools the company is developing for automated patching and red teaming. In red teaming, a tester tries to break into a system as an attacker would, to find its weaknesses.

OpenAI describes Daybreak Defense Factory in more detail on its website. Companies bring findings from their existing security tools and from Daybreak Red Team into the service. AI agents, programs that use a model to act, investigate each finding, trace the vulnerability through the code and develop a patch.

The patches are proposals. Developers review and test them before they are merged into the code, the page says.

Developers can also work on a patch with OpenAI’s Codex: adding context, examining the evidence behind a finding and tuning the patch. They can test changes while the agents keep working in the background.

Defense Factory can also scan new commits, the changes saved to the code, or check the code on a schedule. Developers can hand the investigation of new findings to the agents.

Little was said about Daybreak Red Team beyond its name. Sottiaux’s mention of automated red teaming apparently refers to it.

The Defense Factory page invites users to connect repositories and security findings to the service. It does not give a price or say who can already use the service. A recording of the launch event is available online, Sottiaux said.

What this is based on

1. Daybreak Defense Factory | AI Vulnerability Remediation

   Publisher: openai.com · source publication time unknown · retrieved 13 November 2026 at 05:31 (UTC+02:00) · discovery channel: web

2. Announcing two new cyber products: - Daybreak Defense Factory (https://t.co/OdlGbgPWp8) - Daybreak Red Team (https://t.co/VQTAjDfduz) Both make it much easier to put frontier cyber capabilities to…

   gdb · Publisher: x.com · published 13 November 2026 at 00:16 (UTC+02:00) · retrieved 13 November 2026 at 04:04 (UTC+02:00) · discovery channel: X

3. We announced new tools we're developing for cybersecurity today across automated patching and automated red teaming. Watch the recording here https://t.co/fW371aFjB5

   thsottiaux · Publisher: x.com · published 13 November 2026 at 02:24 (UTC+02:00) · retrieved 13 November 2026 at 04:11 (UTC+02:00) · discovery channel: X

In an announcement published on 12 November 2026, Amp said it is dropping the Dial. The Dial let users choose how demanding a task was from four modes (low, medium, high and ultra), and by default Amp decided which model ran behind each mode.

Amp said billing was the reason for the change. Users can connect their own subscriptions and API keys to Amp, and Amp is then free. Since that became possible, the model has also determined who pays. The medium mode, for example, could use a Claude subscription, a ChatGPT subscription or Amp credits, and users could not easily see which of them was paying.

Amp stresses that users are not choosing just a model. The company builds a harness around each model: the program that steers the agent, gives the model its tools and runs its steps. Each model in the selector has a system prompt and tools optimised for it, and both change with the model.

Switching models mid-thread comes at a cost

A thread is a single conversation with the agent. Under the Dial, a thread stayed in its original mode and model because, Amp said, a model became confused when continuing a conversation started by another model. The company now judges that models recover from this well enough. The new model takes over from the user’s next message, and the conversation history records the switch.

Switching is not free. The prompt cache stores text that has already been processed so that it does not have to be processed again. The cache is specific to each model, however, so the new model reads the whole thread without it, and the first message after a switch costs more. Threads run by a plugin agent stay with that agent.

New threads start by default with Staff Pick, Amp’s own recommendation. It is currently Claude Opus 5.5, and Amp says it will change the recommendation if another model performs better in its benchmarks. The exception is users who had set a default mode for the Dial. In the settings, users can choose to start new threads with the last model selected or with one fixed model.

Dial settings do not carry over

Existing threads continue with the model they started with. Users who had changed a mode’s model in Tune Modes now need to pick that model in the selector. Custom agents added under Build Dial are listed under Other Models.

With default settings, the modes used these models: low ran GLM-5.3-Flash, medium Claude Opus 5.5, high GPT-6 Astra and ultra Claude Fable 5.2, which is now under Other Models. Users can pin models under Settings → Model Selector, and Amp’s Model Selector documentation gives more detailed instructions.

What this is based on

1. Bye Bye, Dial

   Publisher: ampcode.com · published 12 November 2026 at 19:14 (UTC+02:00) · retrieved 13 November 2026 at 05:33 (UTC+02:00) · discovery channel: web

US brokerage Charles Schwab announced its AI connection in a press release on 12 November. The company had $13.33 trillion in client assets in the third quarter, the release said.

MCP (Model Context Protocol) is an open protocol that lets an AI application retrieve information from an external service and use its functions. An MCP server is a program on the service’s side that provides this information and these functions. Clients connect their own large language models to Schwab’s server.

What the AI can do in an account

Once clients have given permission, their AI can retrieve market data and research, analyse the performance of their investments, summarise their portfolio, create alerts and organise investment information, Schwab said.

The AI can prepare, place and manage trades only if the account is thinkorswim-enabled. thinkorswim is Schwab’s suite of trading platforms, with desktop, web and mobile versions.

Schwab said the client stays involved in all of this (‘client in-the-loop’). The release does not say what this involvement means in practice. FinanceX Magazine interprets it as meaning that the client approves every action the AI takes. The magazine said the server has its own authentication and account-identifier controls. Schwab’s release does not contain this information.

The magazine compares the service with Robinhood’s Agentic Trading, which Robinhood launched in beta on 27 May 2026. In that service, agents from Anthropic and OpenAI trade equities autonomously in a separate, separately funded account. FinanceX said independent developers had already published unofficial MCP connections to Schwab accounts.

Options trading before the market opens and after it closes

In the same release, Schwab said clients can now also trade options on about 30 stocks, ETFs and indices on Cboe Options Exchange during extended hours. On weekdays, the morning session runs from 7.30 to 9.25 and the evening session from 16.00 to 16.15 New York time. The release warns that extended-hours trading carries risks, such as lower liquidity, greater price volatility and wider spreads between bid and ask prices.

thinkorswim will also soon get an assistant, thinkorswim Assistant, which will answer plain-language questions about using the platform. AI-based portfolio analyses now cover more stocks and ETFs than before.

The MCP server is in a pilot open to a select group of clients using Claude. Over the coming months, Schwab plans to open access to more clients and support other AI platforms. The release states that Anthropic is not part of the Schwab group. Schwab’s website says thinkorswim is free for its clients, but standard pricing applies to trades, for example $0.65 per options contract.

What this is based on

1. Schwab Announces Platform Enhancements Including MCP Server to Connect Clients’ AI Platforms With Their Schwab Accounts

   Publisher: pressroom.aboutschwab.com · source publication time unknown · retrieved 13 November 2026 at 05:43 (UTC+02:00) · discovery channel: web

2. Schwab MCP Server Links Clients' AI to their Trading Accounts

   Publisher: www.financexmagazine.com · published 12 November 2026 at 16:01 (UTC+02:00) · retrieved 13 November 2026 at 05:42 (UTC+02:00) · discovery channel: web

3. Charles Schwab adds MCP server and extended options trading hours

   Publisher: www.streetinsider.com · published 12 November 2026 at 15:34 (UTC+02:00) · retrieved 13 November 2026 at 05:43 (UTC+02:00) · discovery channel: web

4. ToS Platforms Overview

   Publisher: www.schwab.com · source publication time unknown · retrieved 13 November 2026 at 05:43 (UTC+02:00) · discovery channel: web

Google announced A2UI 1.0 on its developer blog on 11 November. Today’s AI agents mostly answer with long text, the company said. That is a poor way to compare three products, for example, or to edit and approve a detailed plan.

The agent describes the screen, the app draws it

A2UI (Agent-to-User Interface) is an open protocol and a set of libraries. Instead of writing UI code, the agent sends JSON messages that specify a screen’s components and the data bound to them. The app’s renderer converts these messages into the app’s own components. No WebView or HTML view is needed, the project’s GitHub page says.

The agent can use only components that the app developer has registered, and messages are validated against a schema. Google also claims that A2UI binds data on the server, so sensitive values do not reach the model.

Version 1.0 adds bidirectional function calls. In Google’s example, a picker for a flight or other booking checks availability on the spot. Accessibility support and extensions for additional metadata are also new. The Munder Difflin blog reported that the 1.0 specification drops theme colours, so styling comes from the target platform’s own theme.

Jetpack Compose and Gemini still support version 0.9.1

Renderers exist for React, Angular and Lit on the web, Flutter for mobile, web and desktop, and SwiftUI on iOS. The Android Jetpack Compose renderer still supports version 0.9.1; Google said 1.0 support is coming soon. Gemini, too, supports version 0.9.1 for now.

Agent SDKs are available for Python, TypeScript and Dart. They generate a skill and prompts for the model, validate the model’s response and convert it into protocol messages. The web and Flutter renderers also work in MCP Apps, an MCP extension for showing interactive screens inside an AI assistant. As a result, A2UI screens can be shown in Claude and Gemini, for example, Google said.

Google said Atlassian, Alation, Bloomberg and ByteDance are already building on A2UI. Atlassian is bringing it to Jira. Alation shows metric cards and charts built with A2UI in Claude. ByteDance’s Lynx renderer supports version 0.9 and is moving to 1.0. The details about these companies come from Google, and the SD Times report repeats Google’s blog.

On 12 November, the repository’s GitHub releases included no release or tag named 1.0, Munder Difflin reported. The repository has moved to a2ui-project/a2ui and is licensed under Apache 2.0.

A2UI is free, open-source software. Documentation and a quick-start guide are at a2ui.org. For developers upgrading from version 0.9, Google provides a guide that covers the changes.

What this is based on

1. A2UI 1.0: Build delightful, fast, reliable, agentic experiences- Google Developers Blog

   Publisher: developers.googleblog.com · published 11 November 2026 · retrieved 13 November 2026 at 05:49 (UTC+02:00) · discovery channel: web

2. github.com

   Publisher: github.com · source publication time unknown · retrieved 13 November 2026 at 05:49 (UTC+02:00) · discovery channel: GitHub

3. Google releases A2UI 1.0 for generative UI

   Publisher: sdtimes.com · published 11 November 2026 at 19:46 (UTC+02:00) · retrieved 13 November 2026 at 05:49 (UTC+02:00) · discovery channel: web

4. What is A2UI? Google's A2UI 1.0, explained — Munder Difflin Blog

   Publisher: munderdiffl.in · published 12 November 2026 at 02:00 (UTC+02:00) · retrieved 13 November 2026 at 05:49 (UTC+02:00) · discovery channel: web

Will DePue published a report on the Unemployment Labs site on 12 November that relies solely on the models’ published weights. The report finds the Blackwell fingerprint in DeepSeek-V4-Flash and V4-Pro, Moonshot AI’s Kimi K3 and K3.1-Pro, MiniMax-M3 and Ant Group’s Ling-3.1-flash. Earlier models from the same companies carry a fingerprint that matches Nvidia’s older Hopper generation. DePue’s X profile lists earlier work at OpenAI, CellCog reported.

How a chip leaves a fingerprint

Before training, a generator fills the model weights with random numbers, which it produces reproducibly from a seed. PyTorch splits the work across the GPU’s threads. The number of threads depends on how many compute units, or SMs, the chip has: Hopper has 132, B200 and B300 chips have 148, and GB200 and GB300 chips have 152. The same seed therefore places the same numbers in different positions on different chips.

Some weights stay almost unchanged, such as the rows for words not used in training. Weights that change during training can also correlate weakly with their initial values, but the fingerprint does not survive every training run. DePue searched for the seed and thread count that reproduce these traces, then tested the combination on weights not used in the search.

Results and their limits

Under the report’s statistical model, a search of about 6 trillion combinations would by chance produce a match of about 8 sigma at best. The strongest results for the six models range from 49 to 1,385 sigma. On a B300 chip, DeepSeek’s seed reproduced 461 of 469 selected rows in V4-Flash with a correlation above 0.99. A rented GB200 reproduced all 13 preserved rows of Kimi K3. The method was right in all 24 cases where the developer has disclosed the model’s hardware. DePue regards this as a consistency check, not a measure of accuracy.

The fingerprint shows only which hardware generated the initial values. The report says it supports, but does not prove, the conclusion that the models were also trained on Blackwell chips. The weights do not reveal who owned the chips or where they were. Blackwell systems can be rented outside China. The report therefore does not show smuggling or any breach of export controls.

The only other hardware identified that can produce the same fingerprint is the export-controlled AMD MI300X, when its use is limited to 148 or 152 compute units. For V4-Flash, DePue ruled it out on the basis of small differences in the numbers. For the other five models, DePue considers it very unlikely.

Claims on X and scale estimates

DePue called the evidence indisputable in a thread on X, CellCog and The Agent Times reported. The report uses the word only for V4-Flash. CellCog also reported that the thread says researchers at OpenAI and Anthropic confirmed the result. The report does not name them. The code is not public; the report says reviewers can obtain the scripts on request.

The report estimates that each of the five training runs needed on the order of 1,500–3,500 Blackwell GPUs for one and a half to three months. It describes the calculation as uncertain. The calculation assumes that training took place on the same hardware that generated the initial values. Unite.AI and CellCog cite lower figures from the same report; Unite.AI, for example, gives 500–3,000 GPUs per run.

What this is based on

1. Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs

   Publisher: unemploymentlabs.com · published 12 November 2026 · retrieved 13 November 2026 at 05:22 (UTC+02:00) · discovery channel: web

2. Unemployment Labs Reports Blackwell Patterns in Chinese AI Models

   Publisher: www.unite.ai · published 13 November 2026 at 00:53 (UTC+02:00) · retrieved 13 November 2026 at 05:22 (UTC+02:00) · discovery channel: web

3. Report: 6 Chinese AI Models Started on Nvidia Blackwell

   Publisher: cellcog.ai · published 12 November 2026 at 16:00 (UTC+02:00) · retrieved 13 November 2026 at 05:23 (UTC+02:00) · discovery channel: web

4. Report Assigns Blackwell Fingerprints to Six Chinese AI Models

   Publisher: theagenttimes.com · published 13 November 2026 at 01:09 (UTC+02:00) · retrieved 13 November 2026 at 05:23 (UTC+02:00) · discovery channel: web

5. Six Chinese AI Models Show Blackwell GPU Fingerprints

   Publisher: emergent.sh · published 13 November 2026 at 01:45 (UTC+02:00) · retrieved 13 November 2026 at 05:23 (UTC+02:00) · discovery channel: web

Baidu CEO Robin Li unveiled Token Exchange at Baidu World 2026 in Beijing on 12 November. A Baidu press release and Pandaily, which drew on reports from the Chinese outlets IT Home and Yicai, covered the launch.

Baidu promises to check which model responds

Token Exchange works like an exchange: customers use models from several suppliers through a single API. Pandaily reported that Li justified the service by pointing to the many services that have sprung up to resell tokens. Buyers cannot always verify which model is actually responding, Li said.

Baidu says it tests every model on offer before listing it. It compares the model’s fingerprint, an identifier that distinguishes one model from another. It also measures capability, performance and availability, and later carries out spot checks. Suppliers receive a credit rating, and users get a warning if a rating falls. Baidu has not disclosed how the fingerprinting works or how it calculates the ratings.

Baidu says the API supports the OpenAI and Anthropic protocols. Several suppliers can offer the same model, and the service shows each one’s price, first-token latency, throughput and seven-day success rate.

The press release says that after one month the service has more than 100 model resources and handles a daily average of more than 50 million calls and more than 700 billion tokens. The service’s product page gives more than 70 resources, more than 60 million calls and more than 200 billion tokens a day. The figures have not been independently verified.

A reasoning model built for search

Baidu added ERNIE Lightning, which the press release calls ERNIE Light Deep Thinking, to its Qianfan cloud platform. Baidu originally built the model for its search engine to handle complex queries and tool calls quickly. A fast model serves users, while a slower reasoning model evaluates the fast model’s answers and continuously improves it. Baidu Sparse Attention (BSA) compresses cached information about earlier text during generation to cut inference costs.

Baidu says the model runs on a single eight-card Kunlunxin server and cuts inference engine costs by more than 30 per cent, Pandaily reported. No benchmarks have been published.

Qianfan and OpenRouter list ERNIE-Light-DT-1.0, which Baidu says is built on open-source GLM models and research from the Lightning model. Pandaily noted that Baidu has not said whether the two are the same product or disclosed the models’ context windows. Citing Qianfan’s documentation, the AICoder website reported that ERNIE-Light-DT-1.0 has a context window of 1 million tokens and costs 0.008 yuan per 1,000 input tokens and 0.028 yuan per 1,000 output tokens.

Video model and chips

Another new model, ERNIE-Wizshot-ME, makes the person in an image speak in a 4–15-second 1080p video. It accepts scripts of up to 4,000 characters and has a list price of 0.7 yuan per second. Baidu claims its production cost is half the industry norm.

Baidu AI Cloud also presented a 1,024-card Tianchi supernode with fully optical interconnects. The press release says Kunlunxin’s M100 inference chip has shipped and the M300 is expected in 2027.

Enterprise customers and partners can use the models’ APIs on Qianfan. Token Exchange runs at cysuo.com. Users must register with a phone number and verify their identity.

What this is based on

1. Baidu Puts ERNIE Lightning Reasoning Model and ERNIE-Wizshot-ME Expressive Video Model on Qianfan

   Publisher: pandaily.com · published 13 November 2026 at 05:25 (UTC+02:00) · retrieved 13 November 2026 at 05:45 (UTC+02:00) · discovery channel: web

2. Baidu Launches Token Exchange, a Model Marketplace That Fingerprints Models and Rates Suppliers

   Publisher: pandaily.com · published 13 November 2026 at 05:26 (UTC+02:00) · retrieved 13 November 2026 at 05:45 (UTC+02:00) · discovery channel: web

3. Token Exchange-链接模型供应方与消费方的模型交易服务平台-百度智能云

   Publisher: cloud.baidu.cn · published 11 November 2026 at 08:16 (UTC+02:00) · retrieved 13 November 2026 at 05:45 (UTC+02:00) · discovery channel: web

4. Baidu World 2026 Launches Two ERNIE Application Models: a "Lightning Deep Thinking" Model for Low-Cost 1M-Context Reasoning and Wizshot Micro-Expression for Per-Second 1080p Talking-Head Video —…

   Publisher: aicoder.com · published 12 November 2026 at 05:10 (UTC+02:00) · retrieved 13 November 2026 at 05:45 (UTC+02:00) · discovery channel: web

5. Baidu opens ERNIE application models for search tasks and expressive video — Eigen Radar

   Publisher: eigenradar.com · published 12 November 2026 · retrieved 13 November 2026 at 05:46 (UTC+02:00) · discovery channel: web

6. Baidu Unveils Latest ERNIE Models Amid Broader Full-Stack AI Lineup, and Debuts Token Exchange to Power Intelligent Organizations

   Publisher: www.prnewswire.com · published 12 November 2026 at 16:08 (UTC+02:00) · retrieved 13 November 2026 at 05:46 (UTC+02:00) · discovery channel: web

7. Model List - QianfanBaidu | Document - Baidu AI Cloud

   Publisher: intl.cloud.baidu.com · published 7 September 2026 · retrieved 13 November 2026 at 05:45 (UTC+02:00) · discovery channel: web

China’s iFlytek (科大讯飞) announced on 12 November that its SparkAudio 1.0 model is now available to developers through an application programming interface (API). The Chinese technology site IT Home (IT之家) reported the news the same day.

SparkAudio 1.0 is a foundation model for speech: a model trained on a broad dataset and used as the basis for many tasks. It belongs to iFlytek’s Spark model series (讯飞星火). The model accepts speech and text as input and responds in text.

The company says the model can transcribe speech, translate between languages and recognise dialects and ambient sounds. It also separates speakers, analyses emotions and answers complex questions about an audio recording.

iFlytek says the model interprets speech directly, not via text

iFlytek’s case for the model is that audio has usually been processed in a pipeline. Speech is first converted into text, and only then does a large language model interpret it. Text, the company said, keeps only the words; tone of voice, emotion and background sounds are lost. Errors easily accumulate from one stage to the next, and latency builds up when transcription, speaker identification and translation run as separate components.

SparkAudio 1.0 does not merely convert speech into text but interprets it directly, the company said. It recognises speech, ambient sounds and music in one pass, together with their meaning, emotion and situation.

iFlytek says the model supports 99 languages and 202 dialects. The claim is the company’s own. The reports give no accuracy figures and do not list the supported languages, so support for Finnish, for example, must be checked separately.

The company lists customer service, car cockpits, multilingual meetings, patient records and content moderation as use cases. It says the model was trained entirely on domestic, that is Chinese, computing capacity.

Details come from the company; API available on developer platform

An article by the Chinese IT news site IT时代网, published on Tencent News, reports the same details. By its own account, it was compiled from IT Home and other sources, partly with the help of AI, so it does not independently confirm iFlytek’s claims.

The API is available on iFlytek’s developer platform, iFlytek Open Platform (xfyun.cn), and the model can be tried on the company’s website. The reports do not give a price.

What this is based on

1. 科大讯飞星火语音基座大模型 SparkAudio 1.0 上线 API 服务，支持 99 种语言、202 种方言 - IT之家

   Publisher: www.ithome.com · published 12 November 2026 · retrieved 13 November 2026 at 05:46 (UTC+02:00) · discovery channel: web

2. 科大讯飞星火语音大模型SparkAudio上线API服务_腾讯新闻

   Publisher: news.qq.com · published 12 November 2026 · retrieved 13 November 2026 at 05:46 (UTC+02:00) · discovery channel: web

On 11 November, Bloomberg reported on an internal meeting of the AI team at ByteDance, which owns TikTok. The report relies on anonymous sources because the meeting was private. ByteDance did not respond to Bloomberg’s request for comment. The Japan Times also published the Bloomberg story.

Co-founder Zhang Yiming said ByteDance could compete with Anthropic and OpenAI at the cutting edge, the sources said. He cautioned, however, that it would take until at least 2029 or 2030. Zhang now spends much of his time shaping ByteDance’s AI strategy behind the scenes.

ByteDance plans to release an AI agent, a program that handles tasks for the user with the help of a language model, in the coming weeks. The agent resembles Muse, which Meta unveiled in September. Muse is a voice-controlled assistant with an animated character. Meta offers it in its smart glasses and in the upcoming Muse Charm device. ByteDance hopes the agent will help it close the gap with the top US labs in as little as three years, Bloomberg reported.

The agent’s name and features have not been disclosed, and it is not known whether it will have its own device. Heise online reported that the agent is meant to handle everyday tasks such as emails and online orders. Bloomberg’s report does not mention this.

Zhang says training data is orders of magnitude smaller than rivals’

Zhang’s biggest concern is training data, the sources said. He said ByteDance’s Seed unit pre-trains its large language models on datasets that are orders of magnitude smaller than those used by frontier rivals. In pre-training, a model learns language from a large body of text before fine-tuning.

Zhang also reiterated his view that the company should not rely on distillation, that is, training its own model on the outputs of another, larger model. White House officials and Anthropic chief executive Dario Amodei have criticised Alibaba, Moonshot AI and Z.ai, among others, for distilling US models, Bloomberg reported.

Seed is the only major Chinese AI player that has released neither a language model reaching the top of global benchmarks nor an open-weight model, Bloomberg said. Its latest major update, Seed 2.1, came out in June. Even so, the company’s Doubao is China’s most popular ChatGPT-style chatbot. Seed, a unit of about 1,000 researchers and engineers, is led by Wu Yonghui, whom Zhang lured from Google DeepMind in 2025.

ByteDance already leads in video AI

Video generation is the only area of AI in which ByteDance is clearly ahead, Bloomberg reported. The company has trained its Seedance models on its extensive video data and makes money from them. The models have filled the gap OpenAI left when it shut down its Sora service.

Central to Zhang’s ambitions is a plan to spend as much as $70 billion and build Seedance into China’s leading video AI platform, Bloomberg wrote. The Next Web describes the sum as money earmarked for AI infrastructure.

What this is based on

1. 바이트댄스, 오픈AI 추격에 최소 3년 전망…'뮤즈' 유사 AI 에이전트 준비 - AI타임스

   Publisher: www.aitimes.com · published 12 November 2026 at 10:05 (UTC+02:00) · retrieved 13 November 2026 at 05:46 (UTC+02:00) · discovery channel: web

2. Donnerstag: Nvidia-Chef angeblich kurz vor Abschied, KI-Agent auch von ByteDance

   Publisher: www.heise.de · published 12 November 2026 at 07:15 (UTC+02:00) · retrieved 13 November 2026 at 05:46 (UTC+02:00) · discovery channel: web

3. ByteDance prepares Muse-like agent in push to close AI gap

   Publisher: www.japantimes.co.jp · published 12 November 2026 at 03:18 (UTC+02:00) · retrieved 13 November 2026 at 05:47 (UTC+02:00) · discovery channel: web

4. ByteDance reportedly plans a Muse-like AI agent within weeks

   Publisher: thenextweb.com · published 11 November 2026 at 13:31 (UTC+02:00) · retrieved 13 November 2026 at 05:47 (UTC+02:00) · discovery channel: web

On 10 November, OpenAI researchers published a study on the company’s Alignment research blog about how stages of training affect chain-of-thought monitoring. The chain of thought is the text a reasoning model writes as intermediate steps before its answer. The chain usually reveals the model’s intentions, OpenAI said, and the company monitors it as part of its safety work.

The study was prompted by GPT-6 Astra, released in September, whose chain of thought OpenAI said was harder to monitor than that of earlier models. Astra was better than earlier models at evading monitoring mainly in tests where it was prompted to do so.

The UK AI Security Institute (UK AISI) estimated Astra’s time horizon on maths tasks without a chain of thought at 30.9 minutes (median estimate). This means the model succeeds at about half of the tasks that take a human that long. Earlier research had not projected about 7 minutes until the end of 2028. AISI said the estimate may be too high if the tasks have ended up in training data.

Two indirect measures

The researchers did not measure monitorability directly but used two proxies for it, which they themselves consider imperfect. The first is the ability to solve tasks without a chain of thought: the more a model reasons invisibly, the less it needs to write. The second is chain-of-thought controllability, meaning how well the model modifies its chain on request, for example by inserting the word ‘meow’ between all words.

Base model perplexity measures how poorly a model predicts text. As perplexity fell, the ability to solve tasks without a chain of thought improved steadily from GPT-5 to the Astra class. Relative to perplexity, Astra’s result was as expected, even though, measured against time, it outpaced the earlier rate of progress. Controllability was less closely linked to perplexity.

In most tests, architecture made no difference once perplexity was taken into account. In multi-hop planning and serial computation, however, one unnamed architecture performed better than perplexity predicted.

Reinforcement learning clearly increased the controllability of an Astra-class model, even though training only constrained length and graded the final answer. In a smaller experiment, browser and terminal tools, together with grading of their use, increased controllability only slightly. The researchers said the experiments do not fully separate the effects of perplexity and of grading tool use.

Controllability can also pass from one model to another. When the same base model was fine-tuned on responses from two almost equally capable teacher models, the student of the more controllable teacher became more controllable.

In two models, controllability did not predict monitoring evasion

The researchers tested two equally capable models that differed in controllability. The more controllable model was no better at evading monitoring, and in one test it did even worse. Greater controllability alone therefore does not show that models would mislead monitors.

The researchers nevertheless conclude that chain-of-thought monitoring becomes harder as models grow more capable. They expect further progress in base models to require new approaches to monitoring. In the view of the news site The Decoder, developers trade off capability, cost and monitorability, for example when choosing an architecture.

What this is based on

1. What affects chain-of-thought monitorability across the model lifecycle?

   Publisher: alignment.openai.com · published 10 November 2026 · retrieved 13 November 2026 at 05:34 (UTC+02:00) · discovery channel: web

2. OpenAI’s data suggests that frontier models are hiding their thoughts because developers optimize for capability and compute

   Publisher: the-decoder.com · published 12 November 2026 at 18:58 (UTC+02:00) · retrieved 13 November 2026 at 05:34 (UTC+02:00) · discovery channel: web

Perplexity announced two new decision models in a blog post on 12 November. Decider v1.3 is based on Qwen3.8-27B. The smaller Decider Flash v1.3 was trained from Qwen3.5-4B to imitate the larger model’s answers.

The model does not write an answer; it returns probabilities. Users give it text or an image together with questions and their answer options. A question can be a yes-or-no choice, multiple choice or an ordered scale. Perplexity itself uses the model to filter web pages by quality and to check AI agents’ security logs, for example.

How Decider works

Perplexity turns a large language model into a classifier. A decision head replaces next-token prediction and reads the answer at the position of a special marker token. Some of the model’s layers read the input in both directions. In version 1.3, Perplexity isolated the questions from one another so that an answer depends less on what else is asked at the same time.

The model card on Hugging Face says that the model’s software library nevertheless runs each question separately by default. In that case, other questions do not affect the answer. The shared run used in training is cheaper when many questions share the same long input, but in that mode 1.5–2.2 per cent of top answers differ from the default mode, the card says.

The new Locate feature extracts requested spans, such as personal names or personal data, verbatim with their character positions. Perplexity keeps the feature as a separate add-on because merging it into the model weakened ordinary classification answers. Locate works only with text. If a request contains both extraction and classification questions, the model needs a separate compute pass for each, which adds latency.

The company’s own results

The figures are Perplexity’s own measurements. Unite.AI’s report is based on the company’s blog and does not verify them independently.

Perplexity tested search-result reranking on 72 tasks. The metric was nDCG@10, which measures how well the most relevant results rise into the top ten. Decider v1.3 scored a mean of 83.99 points. That is 1.81 points more than Qwen3-Reranker-8B and 2.24 more than TypeSafe AI’s Jev 1.14, which Perplexity names as Decider’s inspiration.

On 63 search-query classification tasks, mean accuracy was 91.6 per cent. That is 6.54 percentage points higher than Perplexity’s current classifiers, but a large language model had labelled the correct answers. In a sample of 1 million real API requests, v1.3 agreed with the language-model judges on 93.98 per cent of questions, v1.2 on 93.68 per cent and Jev 1.14 on 93.35 per cent. The figure measures agreement with another model, not hits against a verified correct answer.

Unite.AI reported that Perplexity put the cost of the same sample at $26.47 for v1.3 and $13.24 for Flash. Perplexity ran Decider on its own hardware. The figures do not show how hardware costs were calculated, and latency was not measured side by side.

Availability and requirements

Perplexity says the models can be downloaded from Hugging Face under the Apache licence and can also be used through OpenRouter and the company’s Decisions API. However, the model card, created on 10 November before the announcement, describes the repository as private, and downloading requires a token with access to it.

The large model needs a CUDA-compatible GPU with room for about 49 GiB of BF16 weights plus working memory. The model card says v1.3 requires its own request format: the prompt format of earlier versions should not be reused because the answers change. Perplexity advises tuning decision thresholds on a small labelled dataset of your own, because the right threshold varies by use case.

What this is based on

1. Perplexity Decider v1.3: Training, eval, and use cases

   Publisher: www.perplexity.ai · published 12 November 2026 · retrieved 13 November 2026 at 05:43 (UTC+02:00) · discovery channel: web

2. perplexity-ai/pplx-decider-v1.3

   Publisher: huggingface.co · source publication time unknown · retrieved 13 November 2026 at 05:43 (UTC+02:00) · discovery channel: web

3. Perplexity Releases Decider V1.3 Models with Text-Span Extraction

   Publisher: www.unite.ai · published 12 November 2026 at 20:45 (UTC+02:00) · retrieved 13 November 2026 at 05:43 (UTC+02:00) · discovery channel: web

Paul Irish of Google described on the Chrome for Developers blog on 12 November how the Chrome team rebuilt the search engine of the Modern Web Guidance skill three times in six months. The skill is an open-source collection of guides, and its stable version was released the same week. It steers coding agents towards new web platform features instead of old workarounds.

The collection holds more than 200 guides on HTML, CSS and browser APIs. At first, the agent sees only the skill’s description of about 205 tokens. On a frontend task, it searches the guides with a CLI tool and receives ten hits, which take up about 530 tokens. It then loads the guide it has chosen in full, about 800 tokens. Irish said MCP tools either waste tokens or agents do not call them reliably, so the team chose a CLI.

The old model did not know the new APIs

The first version took up 507 MB when installed and needed a C++ build that often failed in cloud sandboxes. The second version ran the all-MiniLM-L6-v2 embedding model in JavaScript. An embedding model converts text into a vector of numbers, and texts with similar meanings get vectors that lie close together.

In production, the search failed. MiniLM, trained around 2018–2020, split new API names such as fetchpriority into meaningless fragments. A similarity threshold of 0.38 also rejected relevant results. A search for ‘confirm destructive action’ returned an empty list, and the agent used a window.confirm() call instead of the dialog element.

The team built a new benchmark from agents’ real searches. Telemetry, which can be switched off with the setting DISABLE_TELEMETRY=1, records only the agent’s 2–5-word search phrase. The team added BM25 keyword search alongside semantic search and raised the number of results to ten. The hit rate – how often the expected guide appears among the top results – rose from 80.6 per cent to 94.7 per cent.

Guide embeddings in the cloud, queries from a lookup table

In the third version, Google’s gemini-embedding-001 model computes the guide embeddings at release time, because the guides do not change between releases. Queries are converted into the same vector space without a neural network. The team used Gemini to compute embeddings for 29,165 words and fine-tuned the word vectors on 33,000 real searches. At runtime, a seven-line loop sums the words’ weighted vectors.

The method is designed for short queries, and, according to telemetry, 86.7 per cent of searches were 2–5 words long. When the team tried using the lookup table to embed whole guides, the embeddings lost their ability to discriminate. The dialog guide was then more similar to the search ‘best pizza in new york’ than to the search ‘dialog’.

The table shows the team’s own measurements of the second and third versions.

Source: Chrome for Developers blog, the team’s own measurements. Version 2 hit rate with tuned hybrid search.

The hit rate improved by only 0.7 percentage points over the tuned second version, so the gains are mainly in size and speed. When queries were embedded in the cloud with the Gemini API, the hit rate was 96.9 per cent.

The skill, its search engine and the benchmark are open source. The search tool runs locally without an API key.

What this is based on

1. github.com

   Publisher: github.com · source publication time unknown · retrieved 13 November 2026 at 05:44 (UTC+02:00) · discovery channel: GitHub

2. Local vector search for coding agents: From 507 MB to a 7-line loop | Blog | Chrome for Developers

   Publisher: developer.chrome.com · published 12 November 2026 · retrieved 13 November 2026 at 05:44 (UTC+02:00) · discovery channel: web

The model card for MetaEncoder v1, the model’s user guide, appeared on 12 November in the Hugging Face repository facebook/meta-encoder-v1. The Deploy Log, which lists AI releases, names Meta as the publisher. It cites the same Hugging Face page and offers no information of its own.

The instruction changes the embedding

An embedding is a sequence of numbers that represents the meaning of content. Similar texts or images get sequences that lie close to each other, so they can be searched, compared and clustered. MetaEncoder produces a vector of 6,656 numbers for each input. The input can be text, an image, video frames or a combination of these.

According to the model card, the same input gets a different embedding under a different instruction. An instruction might be, for example, ‘Represent the given image.’ or ‘Find a document image that matches the given query.’ Search queries and the documents being searched usually get different instructions.

The Python code supplied with the model has three main calls: encode produces embeddings, retrieve ranks a collection against a query and decide answers questions about a given situation. The code formats prompts itself in the same way as in training.

Decisions are chosen from options

Decide handles yes/no, multiple-choice and scale questions as retrieval. It embeds each option, and the score depends on how close the option’s embedding is to the question’s embedding. A multiple-choice question can have 2–26 options. Each question is decided independently of the others.

By default, answer probabilities are calculated with the temperature used in training, 0.02. Users with labelled examples can fit the temperature to their own data with the calibrate function. Options can be given descriptions, but the card says the training data for these question types contained only the option names.

Requirements and open questions

The model card gives no benchmark results as text, only an image named benchmarks.png. The card was edited several times within a few hours on the release day. According to the change history, the metadata line ‘license: apache-2.0’ was replaced with other information, and the card text does not mention a licence. Users should therefore check the terms of use in the repository before using the model.

The model processes images at native resolution up to 4,096 tokens. If a video has more than eight frames, the model samples eight at even intervals. It does not truncate inputs, but training sequences were at most 16,384 tokens long.

The model is downloaded from Hugging Face and requires version 5.15 or later of the transformers library. The model card says the model fits on a single 80 GB GPU in bf16 precision. The card also explains how to use the model directly with transformers library classes.

What this is based on

1. facebook/meta-encoder-v1

   Publisher: huggingface.co · published 12 November 2026 · retrieved 13 November 2026 at 05:44 (UTC+02:00) · discovery channel: web

2. MetaEncoder v1 | The Deploy Log

   Publisher: deployedbyai.com · published 12 November 2026 · retrieved 13 November 2026 at 05:45 (UTC+02:00) · discovery channel: web

In an article published on 12 November, Tidningen Balans asked Sweden’s largest financial management software vendors how far their MCP support has progressed. The magazine is published by FAR, the professional body for auditors and accounting firms. The article relies on the vendors’ own answers.

MCP (Model Context Protocol) is an open standard. It lets an AI assistant such as ChatGPT, Claude or Google Gemini use the data and functions of another system. A user can, for example, ask the assistant to fetch a report instead of searching the menus. Balans wrote that what an assistant can do depends on the vendor’s implementation and the user’s access rights.

Spiris and Finago already live

Spiris’s official MCP server has been in open beta since June 2026. Simo Panas, who is responsible for accounting firm customers, said the assistant can create vouchers, process invoices and record payments, among other things. Once a client or an accounting firm consultant has connected the server to their AI tool, the AI can work directly in the books, Panas said. About 1,200 users at about 1,700 companies use the service, the company said. Spiris recommends Claude and ChatGPT, which it has tested, but other tools that support MCP can also connect to the server.

Jan Söderqvist, industry expert at Spiris, expects MCP to produce analyses, forecasts and reports in an instant and to bring the time needed to produce advice close to zero.

Finago launched an MCP solution for its Finago Office system on 15 October 2026, and it is available in Sweden. Country manager Andreas Näslund said the accounting firms that have adopted the service have about 1,500 companies in their client portfolios combined. The company did not disclose the number of firms or active users. A consultant can connect their own AI assistant to the service. Balans’s table states that external AI agents can access the service for approved clients (‘för godkända klienter’). An AI agent can process data from several clients and look for anomalies, for example, Näslund said.

The system checks the permissions in force on each call and logs all data retrieval, the company said. ‘The AI agent never gets broader access than the connected user has in Finago Office,’ Näslund said. Permissions are also client-specific, Näslund added.

Lundify for selected users, Fortnox plans 2027 release

At Björn Lundén’s Lundify, a handful of selected partners and early adopters use the MCP server, said Henry Arousell, the group’s chief product and technology officer. For now, an accounting firm consultant can read and review data for their companies. Support for creating and changing data is coming soon, Arousell said.

Lundify requires a login when the connection is set up, and the consultant has the same permissions as when using Lundify directly, Arousell said. The company plans to open the service to more customers in stages during the rest of 2026 and 2027.

Fortnox does not have its own MCP server. Chief technology officer Magdalena Hallbrandt said the company plans to release an MCP solution in the first half of 2027. Third-party solutions already use Fortnox’s open APIs.

Bokio has no MCP solution. Adam Norlin Persson of Bokio said the company is prioritising the area but did not give a release date.

What this is based on

1. MCP kan bli en revolution för redovisningen – så långt har leverantörerna kommit - Tidningen Balans

   Publisher: tidningenbalans.se · published 12 November 2026 at 16:02 (UTC+02:00) · retrieved 13 November 2026 at 05:39 (UTC+02:00) · discovery channel: web

Harvey, which builds AI tools for lawyers, announced the updates on its blog on 12 November. The latest redlined draft shows the changes since the previous version, the company said, but not why earlier terms were accepted or where the legal department gave ground. Harvey said that information is scattered across emails and old drafts.

Each new version is reviewed against earlier rounds

When the counterparty returns a new version, Contract Intelligence links it to the existing contract. The application adds the version to the overview alongside prior rounds and related agreements. Harvey said the tool takes into account positions the legal department has already accepted when it reviews the new round. This also lets another lawyer continue the review if the original reviewer is away.

Contracts are given statuses: in review, awaiting the counterparty and signed. The company said these let the legal department see in one place what stage each contract has reached. Nobody needs to update a separate spreadsheet.

A signed contract goes into the contract repository, where the terms the legal department accepted are extracted. Harvey’s AI agents for contract review, Contract Review Agents, use them when suggesting changes to the next contract. Users can also ask about their contracts, for example which suppliers objected to security terms and which terms were eventually accepted.

Harvey says this helps contracts close faster. The company gives no figures to support the claim. The features are in early access: the company directs interested customers to contact their account manager or request a demo.

Agents sort lawyers’ email

Blockchain.News reported on 13 November that Harvey has released Harvey Agents for Outlook, agents for the Outlook email program. The agents prioritise emails, draft responses, create tasks and organise correspondence, the publication said. A lawyer reviews and approves each action.

In the article, people with legal backgrounds, some of them working as legal engineers, describe possible uses. They say the agents could compile a daily briefing of meetings and urgent messages, draft calendar entries based on court orders and file messages in Harvey’s Vault document store. These are ideas about how the agents could be used, not measured results. Blockchain.News does not say it tested the tool. Harvey’s blog carries the company’s own post on the same subject, How Legal Engineers Would Use Harvey Agents for Outlook.

Blockchain.News said demos of Harvey Agents for Outlook can be requested through Harvey’s platform.

What this is based on

1. Close Contracts Faster With Every Version in One Place

   Publisher: www.harvey.ai · published 12 November 2026 at 18:00 (UTC+02:00) · retrieved 13 November 2026 at 05:40 (UTC+02:00) · discovery channel: web

2. Harvey Agents for Outlook Streamline Legal Inbox Management

   Publisher: blockchain.news · published 13 November 2026 · retrieved 13 November 2026 at 05:40 (UTC+02:00) · discovery channel: web

The US technology publication The Information reported on the rule on Wednesday 11 November, citing two sources. The rule could be announced as early as this month, according to the report as summarised by Investing.com. However, other agencies are still reviewing it, and it is not final.

The US has already restricted direct exports of the most advanced AI chips to China. Chinese companies have nonetheless been able to rent Nvidia GPU computing from data centres in Singapore and Malaysia, for example, and train models without bringing the chips into China. The arrangement is legal because the Trump administration is not applying the Biden-era AI Diffusion rule, the financial publication 24/7 Wall St. reported. Bloomberg’s Odd Lots podcast said there was reportedly nothing illegal about it.

South Korea’s AI Times, citing The Information, reported that exports of Nvidia’s Blackwell chips and newer generations would require a licence for seven countries: Singapore, Malaysia, Thailand, Indonesia, Brazil, Norway and South Africa. Data centres in these countries would have to tell the Commerce Department who receives the chips and what they are used for. Data centres run by US companies would not need a licence if they block remote access from China and other countries under a US arms embargo. They would still have to report customer details and the amount of computing capacity.

Germany’s heise, citing the same report, gave a different account: Malaysia, Singapore and Thailand would already be on the list of countries requiring a licence. Norway, Brazil and South Africa could be added.

Chinese companies try to reserve capacity in advance

Reuters reported on 12 November, citing three sources, that Chinese cloud providers and AI labs are trying to reserve Nvidia chip capacity outside China. One large cloud company has considered making advance reservations, and some companies have asked data centres about reservations, Le Monde Informatique and AI Times reported, summarising the Reuters story. Smaller AI developers are considering approaching companies in South-East Asia that rent out GPU capacity. AI Times said neither the deals nor their size had been confirmed.

Two of Reuters’ sources also said brokers are building up stocks of export-restricted Nvidia chips on China’s grey market, Le Monde Informatique reported.

The Remote Access Security Act adds to the concern. The bill would give the Commerce Department power to regulate the use of export-restricted technology through cloud services. The House of Representatives passed it in January by 369 votes to 22, and it awaits consideration in the Senate. Nvidia said in July that the law could significantly affect its business, AI Times reported. Le Monde Informatique noted that no general ban on remote access is yet in force.

Nvidia leadership crisis in the background

Nvidia announced on 12 November that chief executive Jensen Huang had resigned with immediate effect. Chief financial officer Colette Kress is serving as interim chief executive. Le Monde Informatique reported that Huang admitted being unfaithful to his wife, and that an internal investigation began after allegations surfaced that the other party, Alice Huang, a former executive of Singapore-based Megaspeed, was linked to Chinese intelligence. Reuters has not been able to verify the allegations, and Beijing denies them.

The Information assessed that the rule may signal Nvidia’s weakened influence in Washington, as the company has opposed tighter regulation of foreign data centres, AI Times reported.

What this is based on

1. NVIDIA Announces CEO Transition

   Publisher: nvidianews.nvidia.com · source publication time unknown · retrieved 13 November 2026 at 05:15 (UTC+02:00) · discovery channel: web

2. U.S., Europe chip stocks dip as Nvidia board reportedly weighs ousting Huang

   Publisher: sg.finance.yahoo.com · published 12 November 2026 at 08:38 (UTC+02:00) · retrieved 13 November 2026 at 05:21 (UTC+02:00) · discovery channel: web

3. U.S. finalizing rule to block China’s remote AI chip access- The Information

   Publisher: finance.yahoo.com · published 12 November 2026 at 06:22 (UTC+02:00) · retrieved 13 November 2026 at 05:22 (UTC+02:00) · discovery channel: web

4. 미국, 중국의 해외 GPU 임대 규제 추진…중국 기업들 확보 서둘러 - AI타임스

   Publisher: www.aitimes.com · published 12 November 2026 at 23:33 (UTC+02:00) · retrieved 13 November 2026 at 05:22 (UTC+02:00) · discovery channel: web

5. Puces Nvidia : les entreprises chinoises d’IA anticipent de nouvelles restrictions américaines

   Publisher: www.lemondeinformatique.fr · published 12 November 2026 at 12:32 (UTC+02:00) · retrieved 13 November 2026 at 05:22 (UTC+02:00) · discovery channel: web

6. Angst vor neuen KI-Chip-Exportkontrollen: China sichert sich Rechenkapazitäten

   Publisher: www.heise.de · published 12 November 2026 at 14:55 (UTC+02:00) · retrieved 13 November 2026 at 05:22 (UTC+02:00) · discovery channel: web

7. Export Controls Were Supposed to Keep AI Chips Out of China. The Rental Loophole Shows Why They Haven’t

   Publisher: finance.yahoo.com · published 12 November 2026 at 18:40 (UTC+02:00) · retrieved 13 November 2026 at 05:22 (UTC+02:00) · discovery channel: web

8. Commerce Department Finalizing Long-Awaited AI Remote Access Rule

   Publisher: www.theinformation.com · published 12 November 2026 at 04:52 (UTC+02:00) · retrieved 13 November 2026 at 05:21 (UTC+02:00) · discovery channel: web

On 9 November, Notion Labs told its customers that a breach at its AI subcontractor Fireworks AI had exposed Notion AI chats. The leak affects less than 1 per cent of Notion workspaces and the investigation is continuing, Notion said in a security notice.

Fireworks AI is a platform that developers use to run open-weight models. Citing a Fireworks statement, SecurityWeek reported that attackers gained access to the company’s research machines on 24 October through a third-party cloud service.

Fireworks notified Notion of the breach on 2 November. Notion initially concluded that only synthetic test data had leaked. By the morning of 3 November, Notion had cut its connections to Fireworks. On 6 November, Fireworks said customer data might also have leaked, and it confirmed the leaked material on 8 November.

Leaked chats came from the Insights feature

Notion used Fireworks for inference on open-weight models and for its Insights feature. Insights shows workspace administrators how members use Notion AI. To run the feature, Notion sent stored chats to Fireworks for classification. Notion has now disabled Insights.

The leaked chats contain everything users wrote to the model and everything the model replied, Notion said. They did not include tool call results. In most workspaces, the chats were not labelled with their author or workspace. Notion nevertheless pointed out that the text itself may contain sensitive information.

Japan’s ITmedia reported that at least three Japanese companies have disclosed a possible leak. Security company SecureNavi said the data may include the names of client companies and contact persons as well as the content of meetings. The companies are still investigating the details.

Notion: Fireworks retained data despite contract

Zero data retention (ZDR) is an agreement under which a provider does not store the inputs sent to an AI model or the model’s responses. ITmedia reported Notion as saying that the leak may also affect enterprise customers with zero data retention agreements.

Notion said its ZDR commitment covers the requests it sends to model providers to generate Notion AI responses. It does not cover features for which Notion must store chats, such as chat history and Insights. For Insights, however, Fireworks had contractually committed to zero data retention, Notion said. Notion’s investigation revealed that Fireworks had retained data that passed through the Fireworks endpoints Notion used.

Fireworks has found no impact on its services where zero data retention is the default, ITmedia reported. Fireworks has not said how the endpoints Notion used relate to those services. Notion says it is now auditing its critical subcontractors.

What customers need to do

Notion has contacted the affected customers. Customers who have not been contacted need take no action for now, Notion said. In most workspaces, members can see their own leaked chats in a desktop browser under Security incident review in the sidebar’s ? menu. Owners can see the affected members in the same place. If the workspace has a verified domain, the owner can get the whole workspace’s leaked content as CSV files.

Notion urges users to change any passwords, API keys and other credentials pasted into Notion AI during the leak. Notion says it has revoked Notion-issued API keys found in the chats and reported keys for other services to their issuers.

What this is based on

1. Fireworks AI incident

   Publisher: www.notion.com · source publication time unknown · retrieved 13 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

2. Cyberattack on Fireworks AI allowed hackers to access Notion customers' chat transcripts | TechCrunch

   Publisher: techcrunch.com · published 12 November 2026 at 17:02 (UTC+02:00) · retrieved 13 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

3. 「AI処理、データ非保持が守られず」――Notionが委託先の問題指摘、漏えいの恐れ巡り

   Publisher: www.itmedia.co.jp · published 13 November 2026 at 12:22 (UTC+02:00) · retrieved 13 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

4. Notionの「AIチャット流出懸念」影響広がる 国内企業が相次ぎ報告 「打ち合わせ内容」も漏えいか

   Publisher: www.itmedia.co.jp · published 12 November 2026 at 16:39 (UTC+02:00) · retrieved 13 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

5. Fireworks AI incident

   Publisher: www.notion.com · published 12 November 2026 · retrieved 13 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

6. Notion Customer Data Compromised in Fireworks AI Hack

   Publisher: www.securityweek.com · published 12 November 2026 at 13:08 (UTC+02:00) · retrieved 13 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

7. Notice about the Fireworks AI incident that may impact some Notion workspaces – Notion Help Center

   Publisher: www.notion.com · source publication time unknown · retrieved 13 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

On Thursday 12 November, Reuters reported on a survey by the IT consulting firm Capgemini of large companies’ AI spending. The survey covered 500 executives in ten countries whose companies have annual revenue of more than $1 billion. Capgemini itself helps companies embed AI into their systems.

Of the respondents, 59 per cent said AI spending over the past 12 months had been significantly above their expectations or budget, by 24 per cent on average. Some 53 per cent of executives are worried about AI costs, up from 45 per cent a year earlier.

Jawlah, which focuses on startup news from the Middle East and North Africa, reports further figures from the Capgemini study. It says 21 per cent of organisations exceeded their budgets by more than 30 per cent. Only 56 per cent are fully confident they can forecast how costs will change as use grows.

Users can download the model weights of an open-weight model and run them on their own servers or on servers they choose. Examples include models from DeepSeek and Mistral. Closed models, such as Claude, ChatGPT and Gemini, are used through the developer’s cloud service.

A cheaper token is not the whole price

Capgemini said tokens for open-weight models were on average 87 per cent cheaper than those for closed models. A token is a chunk of text, and model use is usually priced by the number of tokens.

Capgemini highlights OpenRouter, a service that provides access to models from several developers. In September, open-weight models handled 72 per cent of its traffic but accounted for only 27 per cent of the money spent on use. Jawlah notes that the figures apply to OpenRouter and may not hold for all companies or the market as a whole.

Jawlah also points out that running an open-weight model is not free. A company needs computing capacity and staff to maintain, secure and monitor the models. In the Capgemini study, about half of executives said AI spending was sometimes approved without a sufficient picture of total costs, Jawlah reported. Only about a third can reliably estimate indirect costs such as checking outputs and fixing errors.

Closed models stay for demanding tasks

Half of the respondents plan to use closed models where the complexity of the task justifies the higher price. More than three-fifths of executives expect their organisation to move to a multi-model strategy within the next 12 months, Jawlah said. Under such a strategy, routine tasks go to cheaper models and demanding tasks to more advanced ones.

Reuters links the survey to other news pointing in the same direction. In October, SAP’s chief financial officer said the company would replace some of its most expensive AI models with cheaper ones where performance is maintained. Anthropic launched its IPO roadshow this week. Reuters reported that the company warns in its prospectus that customers may shift to cheaper options, including open-weight models.

What this is based on

1. Major companies turn to open-weight AI models to cut costs | Jawlah

   Publisher: jawlah.co · published 12 November 2026 at 17:37 (UTC+02:00) · retrieved 13 November 2026 at 05:35 (UTC+02:00) · discovery channel: web

2. 研究顯示，多數大型企業計畫增加使用開放權重AI以控制帳單

   Publisher: today.line.me · published 12 November 2026 at 11:01 (UTC+02:00) · retrieved 13 November 2026 at 05:36 (UTC+02:00) · discovery channel: web

3. Most large firms plan higher open-weight AI use to cap bills, study shows - AOL

   Publisher: www.aol.com · published 12 November 2026 at 11:01 (UTC+02:00) · retrieved 13 November 2026 at 05:36 (UTC+02:00) · discovery channel: web

SpaceX announced its plans on Thursday 12 November. The New York Times and Dow Jones were among those that reported the announcement. The figures and promised benefits are the company’s own. The NYT put the total at $74 billion, Dow Jones at more than $74 billion.

Texas campus to generate its own power

The Texas campus will be built on 444 acres (about 180 hectares) of SpaceX-owned land near the Starbase rocket site, the NYT reported. Dow Jones said the company intends to train frontier models there. The sum varies by source: the NYT reported $24 billion, Dow Jones about $24 billion and Anadolu more than $24 billion.

The company said the campus will generate its own power rather than draw it from the public grid. The Next Web reported that SpaceX will build its own gas-fired power plant for the campus and run it itself. The NYT said it is unclear whether the project still needs permits.

The project has already created more than 3,000 construction jobs, and the finished campus is expected to bring more than 500 permanent jobs, the company said.

Dow Jones reported that Gina Hinojosa, the incoming Democratic governor of Texas, has criticised data centres, but analysts mostly expect her to let projects that are already permitted or under way go ahead. SpaceX said: ‘Realizing its full potential will require a stable, competitive, and supportive policy environment that allows projects of this scale to move forward.’

Arizona project awaits county approval

The campus planned for Cochise County, Arizona, would cost $50 billion. The company said it would create about 5,500 construction jobs and 1,000 permanent jobs. Teslanorth reported that the company’s project page describes nine buildings, up to 6 gigawatts of power and 24 gigawatt-hours of battery storage. The company is targeting construction in 2027, but Teslanorth noted that this is not an opening date.

SpaceX said it will work with local electricity co-operatives to use the area’s existing renewable energy, and will cool its equipment with dry cooling instead of groundwater. The company estimates that co-operative members’ electricity bills could fall by as much as 20 per cent. The Next Web said it is unclear whether the Arizona campus will also use gas.

Citing a local newspaper, The Next Web reported that hundreds of residents criticised SpaceX at a meeting last week over water, electricity bills and wildlife. Teslanorth pointed to earlier reports that the county hearing is scheduled for 9 December.

SpaceX rents computing power to Anthropic and Google

SpaceX, which went public this year, rents computing power to other AI companies. In October the company said its AI business generated more than half of its revenue, the NYT reported. Anthropic has committed to paying up to $120.5 billion by 2029, and Google pays $920 million a month.

SpaceX’s capital expenditure rose to $19.6 billion in July–September, nearly three times the figure a year earlier, the NYT reported. In October, chief financial officer Bret Johnsen estimated that capital expenditure for October–December would likely be $35–40 billion.

What this is based on

1. SpaceX unveils over $74B in US supercomputer campus investments

   Publisher: www.xn--anadoluajans-d5b.gov.tr · source publication time unknown · retrieved 13 November 2026 at 05:49 (UTC+02:00) · discovery channel: web

2. SpaceX Plans $74 Billion Investment to Build Data Centers in Texas and Arizona

   Publisher: www.nytimes.com · published 12 November 2026 at 21:29 (UTC+02:00) · retrieved 13 November 2026 at 05:48 (UTC+02:00) · discovery channel: web

3. SpaceX Plans More Than $74 Billion in Texas, Arizona Data-Center Investments

   Publisher: www.morningstar.com · published 12 November 2026 at 19:13 (UTC+02:00) · retrieved 13 November 2026 at 05:48 (UTC+02:00) · discovery channel: web

4. SpaceX plans $74B of AI computing centres in Arizona and Texas

   Publisher: thenextweb.com · published 12 November 2026 at 19:54 (UTC+02:00) · retrieved 13 November 2026 at 05:48 (UTC+02:00) · discovery channel: web

5. SpaceX Unveils $74 Billion Texas and Arizona AI Expansion

   Publisher: teslanorth.com · published 12 November 2026 at 19:56 (UTC+02:00) · retrieved 13 November 2026 at 05:48 (UTC+02:00) · discovery channel: web

6. SpaceX (SPCX) to Spend $74 Billion on Supercomputer Campuses in Texas and Arizona

   Publisher: markets.businessinsider.com · published 12 November 2026 at 20:03 (UTC+02:00) · retrieved 13 November 2026 at 05:49 (UTC+02:00) · discovery channel: web

On 12 November Salesforce released SOV-Transcribe-1, a model that turns audio recordings into text, on Hugging Face. Artificial Analysis (AA), which compares models, published its own measurement on X the same day. The AA-WER Index measures the share of words a model transcribes incorrectly, and a lower figure is better. The model scored 2.3 per cent. AA said this makes it the best open-weight model and seventh of 44 models in a comparison that also includes closed commercial models.

In the chart AA published, the next-best open-weight models both score 2.8 per cent: Mistral’s Voxtral Small and NVIDIA’s Nemotron 3.5 Transcribe. Alibaba’s Fun-Realtime-ASR-preview leads the whole comparison at 1.7 per cent.

In a second post, AA said the model transcribes a minute of audio in about a second on Salesforce’s hosted service.

Architecture and parameters

The model processes the whole recording at once rather than producing text as the speech progresses. The Chinese site hub.baai.ac.cn (智源社区) published an article, compiled by the AI agent AREX, that summarises the model card. That article says the model combines a Whisper large-v3 audio encoder with the Gemma 4 26B A4B language model. The language model is a mixture-of-experts (MoE) model, which uses only some of its parameters to process each token. All the parameters must still be held in memory.

The sources give different parameter counts. Salesforce’s blog puts the active parameters at 4 billion and the total at 26 billion. AA gives 4.6 billion active parameters. The model card lists 4.6 billion active and 26.6 billion in total, including the audio encoder.

The company’s own comparisons

AA has published the model’s word error rate and speed. The other results are Salesforce’s own. The company said the model formats text according to instructions written in plain language. When given keywords such as rare names, the model transcribed them more accurately in the company’s test: the keyword F1 score rose from 71.9 to 96.8.

The table shows the word error rate Salesforce reported in four tests. The company chose both the tests and the models it compared against.

Source: Salesforce’s release. Open ASR and ContextASR rows from AREX’s article. Lower is better.

AREX reported that in the Open ASR test the model trails both rivals in all six languages other than English. Finnish is not among the supported languages listed in the model card. These are English, French, German, Spanish, Italian, Portuguese and Dutch. The longest recording tested lasted 30 minutes.

The weights can be downloaded from Hugging Face under the Apache 2.0 licence. The model card says a single RTX PRO 6000 Blackwell Server Edition GPU, paired with a helper model (SOV-Transcribe-1-DFlash), transcribes up to 1,039 seconds of audio per second. The figure was measured with a November 2026 development version of vLLM that has support for the model built in. Official vLLM does not yet include this support, which has only just been submitted to the vLLM developers. For now, the model runs through a plugin. With the helper model, the limit on concurrent requests must be lowered to 64, because vLLM’s default of 1,024 exhausts the memory and the model does not fit in 96 gigabytes of GPU memory. Salesforce said it will publish a technical report later.

What this is based on

1. Introducing Salesforce SOV-Transcribe-1: #1 Open-Weight Model Built for Real-World Speech

   Publisher: huggingface.co · published 12 November 2026 at 18:43 (UTC+02:00) · retrieved 13 November 2026 at 05:33 (UTC+02:00) · discovery channel: web

2. Salesforce’s SOV-Transcribe-1 Ranks First Among Open-Weight Speech Models | TokenPost

   Publisher: www.tokenpost.com · published 12 November 2026 at 22:06 (UTC+02:00) · retrieved 13 November 2026 at 05:33 (UTC+02:00) · discovery channel: web

3. SOV-Transcribe-1 is the new #1 open weights Speech to Text model, scoring 2.3% on the AA-WER Index with 4.6B active parameters. It ranks #7 of 44 models overall, including proprietary models.…

   ArtificialAnlys · Publisher: x.com · published 12 November 2026 at 18:30 (UTC+02:00) · retrieved 13 November 2026 at 04:14 (UTC+02:00) · discovery channel: X

4. On Salesforce's hosted endpoint, SOV-Transcribe-1 reaches a Speed Factor of 60x, transcribing 1 minute of audio in about 1 second https://t.co/BPjue01o2O

   ArtificialAnlys · Publisher: x.com · published 12 November 2026 at 18:30 (UTC+02:00) · retrieved 13 November 2026 at 04:14 (UTC+02:00) · discovery channel: X

5. Salesforce 开源语音转文本模型 SOV-Transcribe-1 上线 Hugging Face，Artificial Analysis 实测 2.3% 列开放权重第一 - 智源社区

   Publisher: hub.baai.ac.cn · source publication time unknown · retrieved 13 November 2026 at 05:33 (UTC+02:00) · discovery channel: web