Google releases Gemini 4 Carbon for long-horizon agent workflows

#1 𝕏

Google AI released Gemini 4 Carbon, its strongest frontier model, for complex, long-horizon workflows across coding, knowledge work, cybersecurity, and scientific reasoning. It is available through select Google consumer, developer, enterprise, research, and cybersecurity channels, with researchers able to apply for a chance at 6 months of free access.

Also covered by: @Dan Shipper, @Philipp Schmid, @Demis Hassabis, @Josh Woodward

#2 𝕏

Claude Fable 5.2 is now available.

#3 📝 OpenAI News

Introducing Businesses in ChatGPT - OpenAI announced Businesses in ChatGPT in a post dated November 10, 2026. It is intended to let businesses claim and update their information and understand how people engage with them in ChatGPT.

Also covered by: @OpenAI

#4 𝕏

Claude Whiteboards, Claude Sheets, and Claude Data began rolling out on the day of the post, supporting planning on boards, tracker-building in sheets, and storytelling with data.

Also covered by: @Mike Krieger

#5 𝕏

Cursor released Team Projects in beta, providing shared context for teammates and their agents. Agents share plans before starting, flag conflicts while working, and unblock tasks across the team.

#6 𝕏

Qwen announced that Qwen4 is coming, touting deeper knowledge, richer world understanding, and stronger coding and agentic capabilities without providing benchmarks or a release date. Qwen4-Max is in preview on Alibaba’s Token Plan, with early access also available through QwenWork, Qoder, Qwen APP, and QwenStudio.

#7 𝕏

Google DeepMind announced Gemini Robotics 2.5, its latest physical AI, which can learn hundreds of everyday skills, including folding laundry, serving pastries, and autonomously sorting car parts on a factory floor.

Also covered by: @Sundar Pichai

#8 𝕏

Google DeepMind announced that its AlphaProof Code framework mathematically proved Signal’s protocol secure down to the code level—the first such AI verification. Signal’s cryptographic system protects billions of private messages every day.

Also covered by: @Demis Hassabis

#9 𝕏

Mustafa Suleyman shared MAI-Transcribe-2, describing it as the world’s fastest, cheapest, and most accurate transcription model. He claims it is 10x faster than GPT‑Transcribe and 5x faster than Gemini 3.5, at $0.10 per hour of audio.

#10 𝕏

Philipp Schmid announced the launch of MKA (Multimodal Knowledge Agent) with Artificial Analysis, a benchmark for agents working on visually complex files with 200 domain-expert tasks across 10 verticals and 83 occupations. Gemini 4 Carbon ranked #2 on quality, 3 points behind Claude Fable 5.2, at $6.29 per task versus $31.97.

#11 𝕏

Guillermo Rauch recapped that, less than one week after launch, Vercel Security Review had identified 57,897 vulnerabilities—6,029 critical or high—with 14,665 already patched by Agent. He said it can be enabled for any repository, regardless of where it is deployed.

Also covered by: @Guillermo Rauch

#12 𝕏

NVIDIA AI announced that OpenShell is joining @CloudNativeFdn and invited AI agent builders to help shape how agents access data, use credentials, and execute code by contributing to the project.

#13 𝕏

Mistral AI recapped using Voxtral Transcribe 3 to transcribe every main-stage talk at AI Engineer Paris in real time, improving accuracy through context biasing with speaker names and technical terms from the conference agenda.

#14 🐙 herdrdev/herdr

herdrdev/herdr - herdrdev/herdr is an Apache-2.0 Rust app that keeps coding-agent terminals running after clients close or SSH drops, with source and documentation for local and remote sessions, agent status, plugins, a CLI, and a socket API. It has 50,157 GitHub stars and 2,196 weekly trending gains. Worth trying if you need to manage multiple coding agents and machines from one terminal window.

#15 🐙 shadcn/improve

shadcn/improve - **shadcn/improve** is a reusable agent skill that audits codebases and writes self-contained implementation plans for other agents rather than changing source code itself. The MIT-licensed repository includes skill and plugin files plus examples, and has 10,440 GitHub stars with a 333-star daily trending gain. Worth trying if you need structured plans for cheaper executors and use an agent supporting the Agent Skills format.

#16 𝕏

Cognition announced that Claude Mythos 5.2 is available in Devin’s managed security scans, with $100M+ in credits offered to eligible companies to find and fix critical vulnerabilities. Cognition reports a score of 73.5 on its internal cyber benchmark, up from 56.9 for Mythos 5.1; eligibility criteria and conditions apply.

#17 𝕏

Cognition announced that Google’s Gemini 4 Carbon is available in Devin Desktop and CLI. It scores 40.0% on FrontierCode 2.0, up from Gemini 4 Argon’s 26.3%, making it the best-performing Google model on the benchmark.

#18 𝕏

Peter Yang said he pushed a major update to his 3D kart racer after building it all morning with @claudeai Fable, adding three tracks—beach, jungle, and volcano—online multiplayer for up to six friends, six distinct racers, and music. He said it is free to play in a browser.

Also covered by: @Peter Yang

#19 ▶️

I was a Claude hater. Fable 5.2 changed my mind.

How I AI Podcast

#20 𝕏

Julien Chaumond shared how he uses Pi’s Virtual Models to route between Qwen 3.8 27B running locally in @Llama_app and remotely hosted Kimi K3 or GLM 5.3 based on query complexity and machine load, eliminating manual model switching.