Particle.news
Download on the App Store

OpenAI Names Unreleased Astra After Internal Model Produced Ten Machine‑Verified Proofs

OpenAI published Lean certificates that let anyone run automated proof checks yet kept the Astra model private, prompting questions about reproducibility, authorship and safety.

Overview

  • OpenAI said an internal version of its unreleased Astra model generated ten advances in mathematics and theoretical computer science and released manuscripts plus machine-checkable Lean certificates for each result.
  • The company published the Lean files on GitHub so independent proof assistants can verify every logical step, but it did not make the Astra model or the exact discovery process available for outside reproduction.
  • OpenAI reported the token cost to produce the ten results at roughly $2,000 at Sol API rates, a figure it used to underline computational efficiency while noting humans prepared the final manuscripts.
  • Competitors and researchers pushed back: an Anthropic engineer said Claude Fable solved five of the same problems, and mathematicians gave mixed reactions that praised verified steps while urging caution about AI-driven claims.
  • The announcement follows a July internal‑model incident that OpenAI said involved a different deactivated prototype, and the debate now centers on how powerful internal models are tested, who gets access, and how AI contributions should be credited.