Claude formalized Fermat's Last Theorem in Lean — 13M lines, 11 days
First end-to-end computer-checked proof, ~6B tokens, now on GitHub. Plus VISTA sweeps ARC-AGI-3, Codenotch pins usage limits, and Codex takes notes.
Copy markdown
13 million lines, verified line by line
An internal model roughly on par with Claude Fable 5.1 formalized Wiles' proof of Fermat's Last Theorem in Lean 4 — working largely autonomously for 11 days and about 6 billion output tokens, clearing ~29,500 of 30,300 supporting theorems. It's the first end-to-end, machine-checked FLT proof, and the full repo plus a written walkthrough are public: a rare, concrete look at how a coding agent sustains a task over 5x the size of Mathlib.
VISTA wins all 25 public ARC-AGI-3 games
VISTA is an open visual-reasoning harness that gives a multimodal model 'lossless visual memory' and drives it through unfamiliar games straight from raw pixels. Running Claude Opus 5.0 via the Claude Code CLI it clears all 25 public ARC-AGI-3 games at a perfect 100 human-action-efficiency (98.3 with GPT-5.6 Sol via the Codex CLI) — a harness you can point at your own pixel-based or long-horizon agent tasks.
Pin your Claude, Cursor, and Codex limits to the notch
Codenotch is a free, MIT-licensed macOS app that pins live usage limits for Claude Code, Cursor, Codex, and Antigravity to your screen edge — session percentages, reset times, and 429-aware polling. The caveat is honest: it reads unofficial, unstable endpoints, so expect breakage whenever a provider shifts them.
Codex learns to keep notes across context windows
OpenAI shipped experimental context management for Codex: the agent keeps its own notes across context windows and can search back through earlier messages and tool outputs instead of re-compressing everything — while preserving prompt caching. Turn it on with features.context_management.experimental_mode = true (Plus/Pro now, default rollout planned).
Elsewhere: GitHub restores star-history tracking with a safe API
GitHub added a REST endpoint that returns a repo's historical star counts with timestamps but no stargazer identities — reviving growth-tracking that broke when stargazer lists were locked to admins earlier this year. If you build trending boards, README badges, or dashboards, you can wire star history back in today.