GitHub radar
See how OpenAI's unreleased model tackles real math research
OpenAI published 721 math manuscripts on GitHub, written by an unreleased internal model — a raw snapshot of how far AI reasoning has gotten on open research problems.
A public GitHub repo with 721 math manuscripts (proof write-ups) generated by an unnamed OpenAI model. About 42% of the top-line results are checked line-by-line by a tool called Lean; the rest may still contain errors.
When it helps
Useful if you want to see how an AI model reasons through hard math problems and where it still falls short. Skip it if you need a finished tool or peer-reviewed papers — these are raw drafts, not polished research.
Pros
- Shows the real current math level of an unreleased OpenAI model
- Includes short reasoning summaries for 10 selected results
- 42% of top-line results are formally checked with Lean
- Free and open for anyone to browse
Cons
- Not a usable product, just raw research data
- Most results lack formal verification and may have errors
- Needs some math background to make sense of the content
How to set it up — step by step
- 1Open the repo at https://github.com/openai/math and browse the 373 result families.
- 2Pick a topic or manuscript you're curious about using the overview or manuscript map.
- 3Send your agent the text from the block below to get a plain explanation of the problem and approach.
- 4Check whether your chosen result has a Lean formalization — that marks it as more rigorously verified.
- 5Look at the ready-made reasoning summaries (reasoning_traces) if your topic is one of the 10 listed examples.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Open https://github.com/openai/math, pick one manuscript or result family on a topic that interests me, and explain in plain terms what the problem was and how OpenAI's model approached it.
▌ More finds