OpenAI · Oct 7, 2026 · 12:27 PM

OpenAI dumps 722 math manuscripts from an unreleased model into a public GitHub repo, many Lean-checked

Released Oct 6 in openai/math, the papers form 372 result families drawn from about 4,000 problems, averaging roughly three hours of ChatGPT Pro thinking each, with Lean formalizations for many (not all) proofs and 10 abridged reasoning summaries. The IAS advisory group asked for the model, prompts, and per-result compute to be public, but OpenAI is sharing only averages and no prompts, and the model itself stays internal for now.

Read at openai.com

More news

View all

Codex Day 3: GPT-6 lands in ChatGPT's Chat tab and every paid account gets a banked usage reset

Codex lead Tibo's Day 3 post (Oct 7) headlines GPT 6 in ChatGPT Chat, with GPT 6 Sol for paid tiers now and GPT 6 Luna for Free and Go users from Oct 8, while the models in Work and Codex stay unchanged. He also said Codex and ChatGPT Work hit a new high of 40M active users and that a banked reset is loading into every paid account, so you can save it for when you actually hit your limit…

X / @thsottiaux (Codex lead) 21 minutes ago

Codync turns Claude Code, Codex, and 40 more coding agents into bots you message from your phone

Shown on HN Oct 7, the MIT licensed project runs one Rust binary, codync host , on your Mac, Windows, or Linux machine that drives any agent in the ACP registry with the plans you're already signed into, and you chat with each bot (its own folder, memory, and approval rules) from an iPhone app, desktop app, or terminal. Remote access goes through a free end to end encrypted Cloudflare relay, and bots can ask each other for help, like a Claude bot requesting a Codex review…

Show HN / GitHub 1 hour ago

Copilot CLI's /model picker now finds your local Ollama models and switches to them mid-session

Starting in Copilot CLI 1.0.94 0 (announced Oct 7), /model lists tool calling, streaming capable models from a running Ollama instance next to the cloud ones, and you can add one for the current session without restarting. Picking a local model doesn't turn on offline mode or disable telemetry, so set COPILOT OFFLINE=true if you want that…

GitHub Changelog 2 hours ago

Liquid AI open-sources d1-3B, a decision model that answers in 8 ms on an RTX 4090 with zero output tokens

Released Oct 7 on Hugging Face alongside the experimental d1 omni 600M (text plus image or audio), d1 3B returns a typed answer in one forward pass instead of generating text, scoring 48.57 on the Decision Index v0.2.1 public split, which Liquid says matches Decider 35B A3B at 12x smaller. It has day one llama.cpp support and runs a question in 16 to 50 ms on Jetson boards, so it's a cheap local option for gating, routing, or moderation steps in an agent loop…

Liquid AI Blog 3 hours ago

Google ships an official agent skill and gcloud commands so coding agents read live Google docs instead of scraping

Announced Oct 7, npx skills add google/skills skill retrieving developer knowledge teaches Claude Code, Cursor, Copilot, or Antigravity to pull fresh Markdown docs for Google Cloud, Firebase, Android, and more through the Developer Knowledge MCP server, falling back to curl on the REST API. There's also a new gcloud developer knowledge surface (try piping an error trace into answer query ) plus client libraries in seven languages…

Google Developers Blog 4 hours ago

Spotted something we missed? Start a thread.