Thu, Nov 12, 2026
Anthropic launches Claude Fable 5.2 and Mythos 5.2
Anthropic launches Claude Fable 5.2 and Mythos 5.2, DeepMind verifies security properties of Signal’s compiled protocol code, and Gemini 4 Carbon reaches general availability. Fable’s thousand-agent Fermat experiment used a shared SQL database. Even a famous theorem, it seems, eventually needs a project tracker.
Claude Fable 5.2 and Mythos 5.2 launch
Anthropic launches Claude Fable 5.2 for all users, with cybersecurity and biology safeguards; Mythos 5.2 shares its underlying model but requires organizational vetting. Fable designed three working early device prototypes, and an early version powered a 1,000-agent toy Fermat experiment completed in 28.7 hours. Fable is also GA in GitHub Copilot, rolling out gradually with prompt and output retention by default for safety checks, not training.
AlphaProof Code verifies Signal protocol code
Google DeepMind’s AlphaProof Code proved key security properties of Signal’s compiled protocol implementation, including message protection before and after temporary device compromise. Human experts defined the goals; AI agents generated Lean proofs during a four-week verification effort, versus years for expert teams historically. DeepMind is open-sourcing the verification artifacts.
Also
Gemini 4 Carbon reaches general availability — Gemini 4 Carbon is now GA, with introductory API pricing of $2 per million input tokens and $10 per million output tokens through November 30, 2026. Paid Copilot access is also GA, with a gradual rollout.
AWS Emulator reaches general availability — AWS Emulator lets the CLI, SDKs, and agents call AWS APIs locally. It does not reproduce IAM policy evaluation or cross-account and cross-Region behavior.
Meta launches Muse Ember — Meta’s Muse Ember pairs a million-token context with pricing of $0.10 per million input tokens and $0.50 per million output tokens for recurring agent tasks.
Flutter 3.50 makes native interop stable — Flutter 3.50 stabilizes direct Dart calls into Objective-C, Java, and Kotlin, avoiding platform-channel messaging. Dart 3.14 adds automated SDK migrations with dart migrate.
Git patches six vulnerabilities — Git 2.56.1 and maintenance backports fix six vulnerabilities affecting clients and merge services, including attacker-controlled code execution during recursive clones.
ExecuTorch brings PyTorch models to browser GPUs — ExecuTorch’s early-development WebGPU backend runs PyTorch models locally in browsers without format conversion; the same exported artifact also runs on Metal, Vulkan, and D3D12.
Cursor Projects becomes multiplayer in beta — Cursor is rolling out multiplayer Projects in beta, giving teammates’ agents shared context. An internal on-call handoff carried over incident details and work in flight.
Gemini Robotics 2.5 reaches 96% kitting success — Gemini Robotics 2.5 is rolling out to partners; an Apptronik–Mercedes-Benz proof of concept reached 96% automotive kitting success with Apollo 2.
Google launches Gemini Decision and Decisions API — Google’s new multimodal Gemini Decision model returns probabilities, choices, and scores without generating output tokens; it is also available through Vercel AI Gateway.
AlphaGenome 2 adds cell-specific DNA predictions — AlphaGenome 2 predicts DNA variant effects for a chosen tissue, cell type, or state. API access covers non-commercial research, with commercial research in private preview.
Node.js patches HTTP/2 and TLS flaws — Node.js 26.13.1 fixes high-severity HTTP/2 and TLS flaws, including a remotely reachable TLS server crash. Security releases also cover 24.22.3 and 22.23.5.
Cloudflare chooses QuePaxa for Meerkat — Cloudflare’s experimental Meerkat uses QuePaxa to keep global consensus moving through leader failures; evaluations across nearly all continents never completely stalled.
HOPE tests leaner mixture-of-experts models — HOPE chooses experts to prune by modeling how they cooperate; at 50% pruning, one Qwen model slightly outscored its unpruned version on SWE-Bench Pro.
OpenAI deploys Control Room — OpenAI’s Control Room can halt further progress on research inference workloads within 35 seconds of a stop instruction.
Businesses in ChatGPT launches in beta — OpenAI’s Businesses in ChatGPT beta lets verified businesses add product feeds with pricing and availability.
Releases
- pnpm 12.19.0funding reports, builds inminor
nodeLinker.type: loaded, JSON lockfiles, 29% faster macOS installs - oxlint 1.93.0minor
prefer-catchsuggestions,exhaustive-depsremoval suggestions, unusedoxlint-disablecleanup with--fix
Have a good Thursday. Reply with anything we missed or a tool you’d like covered. If a colleague is still reading the long versions, send them this edition.
