ANG⁺ #4 · Two Giant Open-Weight Models, +5 More

Welcome to the fourth ANG⁺ Briefing. Briefing #3 (November 3) covered the big numbers — an IPO filing, earnings, and regulation. This edition is about tools you can download, switch on, and check yourself. We picked what readers can actually act on this week from late October through November 4. Everything is built on first-party material: Hugging Face model cards and license texts, plus official company posts and developer changelogs.

1. Two open-weight giants — Kimi K3.1 Pro (2.8T) lands on Hugging Face, Mistral Large 4 (1 trillion) releases weights

The full weights of Kimi K3.1 Pro, which China's Moonshot AI announced on October 24 and promised "by Nov 4," are now up on Hugging Face at moonshotai/Kimi-K3.1-Pro. A week earlier, on October 29 (UTC), France's Mistral released the weights of its improved Mistral Large 4.

• Kimi K3.1 Pro: a mixture-of-experts (MoE) model with 2.8T total parameters and 104B activated per token, with a 1,048,576-token context. The weights are MXFP4, quantized during training, and split into 96 files. In the company's table it scores 65.7 on Terminal-Bench 4.0 — 0.7 points behind Claude Opus 5.5 (66.4) and ahead of GPT-6 Astra (58.2). On an internal factual-accuracy stress test, its claim-level hallucination rate fell from 23.4% for K3 to 8.3%.
• A capability left out on purpose: for the open-weight release, Moonshot says it identified and removed cyber-related data from training. A checkpoint without that removal solved 90.5% of its internal exploit benchmark (xdev-bench-hard); the released model solves 81.0%. The stated principle: do not release open weights that would extend the absolute frontier of cyber capability.
• Mistral Large 4: a 1-trillion-parameter multimodal model trained from scratch on European infrastructure. Three more weeks of reinforcement learning after the preview lifted Terminal-Bench 4.0 by 13 percentage points to 41%, which Mistral says leads open-weight models. It scores 48% on the Artificial Analysis Coding Agent Index, tied with GLM-5.3, while using 47% fewer tokens per task on average. It has three reasoning modes (Low, Medium, High by default), and the API costs $0.70 input, $0.07 cached input, and $2.10 output per million tokens.
• What the licenses share: both are modified MIT licenses — you can download, modify, and use the models commercially. The catch is selling the model as a service (MaaS). Mistral requires a separate commercial license for any MaaS business; Kimi requires a separate agreement once MaaS revenue exceeds $20 million over any 12 consecutive months. Kimi also requires products with more than 100 million monthly active users or more than $20 million in monthly revenue to display "Kimi K3.1-Pro" prominently.

Our take: Open weights are now within a point of the top closed model, and both companies picked the same kind of license — one that only blocks reselling the model itself. But 2.8T is not something an individual runs at home. For most readers, the real choice is not the weights but the price war among the hosts serving them.

Source: Kimi K3.1 Pro model card (Hugging Face) · Kimi K3.1 Pro License (Hugging Face) · Introducing Kimi K3.1 Pro (Moonshot AI) · Mistral Large 4 open weights (Mistral AI)

2. Models that operate screens got cheaper — OpenAI GPT-6.1 Terra, Gemini API Computer Use goes GA

The cost of running "computer use" agents — models that look at a screen, click, and type — just came down a step.

• GPT-6.1 Terra (October 26): standard API pricing is $1 input, $0.05 cached input, and $5 output per million tokens, 50% below GPT-5.6 Terra and GPT-6.1 Sol. On Agents' Last Exam, a professional-work benchmark, it scores 15 percentage points higher than GPT-5.6 Terra at roughly one third of the cost at medium reasoning; at Max, OpenAI says its cost per task is about one ninth of Claude Opus 5.5's. It solves 50.3% of Terminal-Bench 4.0 (GPT-5.6 Terra: 21.5%), and on an internal set of error-inducing prompts its error rate at low reasoning fell from 53.6% to 16.7%.
• Safety tier: it reached the cybersecurity "Critical" threshold in OpenAI's Preparedness Framework and keeps the safeguards built for GPT-6 Astra. It is available in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users, and in the API as gpt-6.1-terra.
• Gemini API (October 30): Google moved the Computer Use tool to general availability (GA) across both the Interactions and GenerateContent APIs, four months after its public preview in June.

Our take: The biggest obstacle to handing agents "a person looking at a screen" work — reading invoices and keying them into accounting software — has been cost. With two companies shipping lower prices and full support in the same week, any automation you shelved is worth re-quoting now.

Source: Introducing GPT-6.1 Terra (OpenAI) · GPT-6.1 Terra model page (OpenAI API docs) · Gemini API changelog, October 30, 2026 (Google)

3. A fake Claude Code extension in Cursor's marketplace planted an SSH backdoor — check with one line

On October 29 (U.S. Pacific time), Cursor disclosed that Anthropic.claude-code 2.1.312 and 2.1.313 (Windows and macOS) in its VS Code extension marketplace were not developed by Anthropic. They were available from about 2:27am to 1:35pm PT that day.

• What it did: it tried to add an attacker's SSH key to authorized_keys, turn on the local SSH service, and open a reverse SSH tunnel. It persisted through a LaunchAgent on macOS or a scheduled task on Windows, retrying the tunnel roughly every 60 seconds. On most machines without admin rights, enabling SSH failed and the attacker could not log in — but Cursor says the key was added, persistence was installed, and host details were sent.
• What Cursor did: it rotated its keys, rebuilt the marketplace and brought it back online, and updated its indicators of compromise (IOCs) on October 30. It warns that uninstalling the extension may not be sufficient, because standalone persistence scripts keep running.
• How to check (macOS and Linux): in a terminal, run grep -qs svc-telemetry ~/.ssh/authorized_keys && echo "svc-telemetry key found" || echo "No key found; machine not impacted". If the key is found, run the cleanup script from the notice (on Windows, in PowerShell as Administrator) and consider rebuilding the machine.

Our take: If you installed or updated the Claude Code extension in Cursor during that window, run that one line first. Agent tools can reach your terminal and your keys, so a single extension is effectively a key to your servers.

Source: Security notice — malicious version of Anthropic.claude-code (Cursor)

4. Anthropic's Claude Duet — an agent app that speaks first (research preview)

On October 21, Anthropic opened a research preview of Claude Duet, an app that brings up what needs your attention before you ask, keeps working while you are away, and checks with you before acting on your behalf.

• How it works: it runs as a standalone app on iOS and Android and on claude.ai. Each morning it sends a short brief on the most important items across your connected apps; during the day it watches those apps and messages you only when something needs you. You can hand it anything from scheduling meetings to building dashboards, financial forecasts, and web-app prototypes.
• Who gets it first: US Pro, Max, and Ultra subscribers, by waitlist. It is built around Claude Opus 5.5. During the preview it shares your plan's usage limits with the rest of Claude, and its background checks count toward them.
• Sensitive information: following the Claude Managed Agents design, it never directly sees your passwords or sign-in tokens, and actions are approved by the auto mode classifier. A new dedicated classifier and safety system prompt reject requests to share the strictest tier of information.

Our take: A chatbot answers when asked; Duet comes to you. That makes "what do I connect it to" the real settings screen. Before it opens outside the US, decide which parts of your email and calendar you would hand over.

Source: Introducing Claude Duet (Claude)

5. Anthropic donates the Agent Skills standard to the Linux Foundation's AAIF — the second after MCP

On October 22, Anthropic donated Agent Skills — the standard that teaches agents specialized knowledge and workflows — to the Agentic AI Foundation (AAIF) under the Linux Foundation. Anthropic co-founded the AAIF with Block and OpenAI and had already donated the Model Context Protocol (MCP) there.

• How far it spread: since the standard was published in December 2025, more than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot, and VS Code, have integrated it. So far 1 million skills have been published and nearly 280 million installed.
• What is handed over: the Agent Skills name and logo, the GitHub organization, and the agentskills.io domain. Engineers from Anthropic, Google, OpenAI, and others will maintain the standard as individuals. You can keep using it and contributing without joining the AAIF.
• What comes next: Anthropic proposed starting with versioned releases, a formal proposal process, and security guidance for people who write and install skills.

Our take: A skill is just a folder and one SKILL.md. With the standard out of a single company's hands, turning your prompts and procedures into skills now means you can carry them across Claude, ChatGPT, or Copilot. But as item 3 shows, installing someone else's skill is now a supply-chain check, too.

Source: Donating Agent Skills to the Agentic AI Foundation (Anthropic)

6. New this week, and dates for your calendar

Small changes from November 2–4, plus the dates within a month when you will need to change code.

• OpenAI to acquire Pierre (November 4): the company behind the Trees and Diffs libraries that the Codex app uses for browsing files and viewing changes, and behind Code Storage, a Git-based code storage service. After the new Git backend Pando we covered in Briefing #3, this adds the code-review side. Pierre's open-source projects stay open source, and the two companies remain separate until closing, which requires regulatory approvals.
• Claude text watermark (November 3): text generated by Claude Mythos 5, Opus 4.7, and Sonnet 4.6 carries Anthropic's text watermark. On Sonnet 4.6, setting temperature, top_p, or top_k to a non-default value might drop it.
• Claude API message threads beta (November 2): the server keeps the conversation, so each later request sends only the new messages. A thread expires 5 minutes after its last request and is deleted no more than 1 hour after its last use.
• Gemini voice remixing (November 4): modify an existing voice with natural-language instructions and save it as a new voice (gemini-3.8-flash-tts).
• November 30: Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) retires from the API, and fast mode for Claude Opus 4.8 is removed.
• December 1: gemini-2.5-pro automatically routes to gemini-3.8-flash and gemini-2.5-flash to gemini-3.5-flash-lite, billed at the new models' rates. Google lists breaking changes, including function calling and thinking_level, and asks developers to update model strings before then.

Our take: Automatic routing sounds kind, but price and behavior change together. Rather than finding out on December 1 from your bill and error logs, search your code for model names this week.

Source: OpenAI to acquire Pierre Computer Company, Inc. (OpenAI) · Claude Platform release notes (Anthropic) · Gemini API changelog (Google)