On — Nov 11

Google releases Gemini 4 Carbon, its strongest model, for long-horizon software engineering, cybersecurity, and science work. It is in the Gemini API, Google AI Studio, Gemini for business, and for Ultra subscribers. Cursor reports it as the top model on CursorBench at 60.4% at high effort.

Why it spiked. It arrives a few weeks after Argon and a day after Anthropic's Fable 5.2, and model identifiers for Flash and Flash-Lite versions are already circulating (unconfirmed).

• @sundarpichai Announces the rollout for long-horizon software engineering, cybersecurity, and scientific work through the Gemini API and Ultra.
• @cursor_ai Adds Carbon to Cursor and reports 60.4% on CursorBench at high effort.
• @demishassabis Shows Carbon reverse engineering his original Theme Park binaries, adding a ride, and porting it to the web in one shot.

Baidu releases ERNIE 5.5 Flash Preview, a mixture-of-experts model with 280B total and 14B active parameters and a 1M-token context. The API is live on Baidu Qianfan at over 200 tokens per second per user, and Baidu says weights will be open-sourced by November 25.

Why it spiked. Baidu's demos show the model building an Unreal Engine game and Blender scenes end to end.

• @Baidu_ERNIE Lists 280B total / 14B active, 1M context, over 200 tokens/s on Qianfan, and open weights by November 25.
• @ShuohuanWang Shows the model building a wuxia game in Unreal Engine, an airport in Blender, and a short film with its own music.
• @teortaxesTex Calls the scores reasonable with the usual caveats and says Baidu still trails the leading Chinese labs.

Cursor rebuilds Projects for teams, in beta. Each teammate runs their own agents against a shared task board and shared context, and agents share plans, flag conflicts, and ask other teammates' agents for what they need.

Why it spiked. Cursor engineers report internal numbers: one project ran 62 concurrent agents across 11 people and merged 972 PRs in 15 days.

• @cursor_ai Says agents share plans before starting, flag conflicts while working, and unblock each other's tasks; launching in beta.
• @mattyp Explains that a teammate's agent can ask another teammate's agent directly for information.
• @aryamankhawow Reports 62 concurrent agents across 11 teammates at peak and 972 merged PRs over 15 days.

Chrome, with Firefox, MDN, and Microsoft Edge, releases Modern Web Guidance 1.0, a set of vetted skills that steer coding agents toward native HTML and CSS features instead of custom JavaScript. It covers performance, accessibility, and security, installs with npx modern-web-guidance@latest install, and works with Antigravity, Cursor, GitHub Copilot, WebStorm, Codex, and others.

Why it spiked. All four major browser and docs teams back it, and Google featured it on its homepage.

• @ChromiumDev Announces the release with Firefox, MDN, and Edge to stop agents writing JavaScript for things HTML and CSS do natively.
• @sarah_edo Describes it as expert-vetted skills for perf, accessibility, and security, with the one-command installer.
• @LeaVerou Notes the project is featured on the Google homepage and asks for feedback.

In its roadshow video, Anthropic says a Claude Mythos model got Claude's prompt processing running on AMD MI355X GPUs in about three days with almost no human help. Anthropic says this work used to take dozens of engineers six to twelve months.

Why it spiked. Developers read it as evidence that model-written kernel code reduces dependence on NVIDIA's CUDA stack.

• @AndrewCurran_ Quotes the roadshow on starting with AMD GPUs this year and the old cost of bringing up a new chip platform.
• @InvestLikeBest Says Mythos Preview did it in about three days versus an estimated six to twelve months, and quotes the CFO on using three chip platforms.
• @AndrewCurran_ Quotes Anthropic: revenue up more than 45x over six quarters while compute costs grew 10x.

The Institute of Foundation Models at MBZUAI releases PAN 2, a world model that predicts the next state of a scene after each action. Weights, training code, and inference code are under Apache 2.0.

Why it spiked. Few world models ship with training code and a permissive license.

• @IFM_AI Introduces PAN 2 as an open foundation world model that predicts what happens after each action.
• @ericxing Says it uses a Generative Latent Prediction architecture and is evaluated on six dimensions of physical capability.
• @_Melis_143 Points out that weights, training code, and inference code are all Apache 2.0.

OpenAI engineer Tibo (@thsottiaux) posts that Thursday "is feeling codexy" and floats a banked usage reset. Accounts tie this to an unannounced GPT-6.2 release; OpenAI has announced nothing.

Why it spiked. Codex users want their usage limits restored, and Google and Anthropic both released models this week.

• @thsottiaux Says Thursday is feeling codexy and may be a good time for a banked reset.
• @thsottiaux Notes it has been a while without usage resets.
• @kimmonismus Speculates GPT-6.2 ships Thursday; unconfirmed.