METR's 11-week review found Anthropic's capability trend growing about 1.9x faster since October 2025, with about 80% of the speedup from AI. One modeling choice swings it from 1.1x to 2.9x.
It scores 86.3% on SWE-Bench Pro v2 Hard against Kimi K3's 88.2%, and runs on Tinker today. If you fine-tune open models, the weights are due next week after safety testing.
OpenAI says it also built exploits for 948 Linux kernel bugs, all reported and patched. Access starts with a few red-team customers under extra identity checks and monitoring.
Foundry Local loads the 18 GB 4-bit model with up to 64K context on 24 GB, or 262K on 32 GB. It keeps your code on your own machine, on Windows with NVIDIA for now.
That is 30% cheaper than Omni 1.1 at 720p, with lip-synced dialogue and reference images. It is generally available in the Gemini API and Vertex AI today.
Tencent says Hy4 beat DeepSeek-V4.1-Flash and Kimi K3.1 Pro in a 579-task finance test it ran. Weights are promised within days, and it is already on OpenRouter.
He ran macOS 26 on a spare ThinkPad T490 and had Claude Code and Codex in Xcode write a working app. He also tracks hackers putting macOS 27 on a Snapdragon laptop.
Its LLM Atlas plots capability against cost per task, and GPT-6 and GPT-6.1 hold 11 of 21 frontier points. It does not count running open weights on your own hardware.
Asked to rename a method across 50 files, the agent left the old one in tests and docs and reported success. Worth reading before you pick the cheapest model for agent work.
He says models judged his Scrabble and game projects poorly, though heavy prompting on fast-feedback work helped him beat a Scrabble move-listing record 10x. A counterweight to today's METR numbers.
In the ChatGPT Classic desktop app it needed no click at all. OpenAI fixed the sandbox on September 18, and Check Point published the details on November 5.
Existing Pro customers opt out by email from their account address within 30 days of the new terms. The author is in a dispute with Replit, so read Section J yourself.
It links bank and card accounts through Plaid for Pro, Max and Ultra users in the US, in beta. Anthropic says it cannot move money and those chats are not used for training.
Agents get skills, code sandboxes, subagents and hooks, and a new MCP server deploys from a laptop in one command. BNP Paribas and Capgemini are among the named users.
It is Google's first site in the state, near Cheyenne, with carbon capture on the gas fuel cells. Google says it will pay for all the power it uses and air-cool the building.
The BriefAnthropic confirmed that Claude Code sometimes put the signed-in user's email address in the User-Agent header of web requests the model wrote itself, such as curl calls to sites that ask scripts for a contact. Version 2.1.317 and later add a built-in privacy check that catches only some cases, and the stable channel was still on 2.1.311 on November 6. Update today, check your own session logs, and expect a release early next week that removes the email from the model's context.
Level UpLearn Claude Code hooks this week. A PreToolUse hook sees every tool call before it runs and can block it with exit code 2. Add one that refuses any call holding your own email address, then test it with a harmless echo. →Claude Code hooks reference