Critics call OpenAI’s Astra math claims ‘oversold’

OpenAI announced on August 1 that its unreleased Astra model family produced solutions to 10 long-standing open problems across mathematics and theoretical computer science, releasing a 249-page manuscript and fully verified Lean 4 proof certificates on GitHub. The results span fields including high-dimensional geometry, coding theory, group theory, quantum complexity, and extremal combinatorics, with the company reporting the total compute cost at roughly $2,000 at GPT-5.6 Sol API rates.ndtv+2

What Astra Solved

According to OpenAI's blog post, Astra generated advances in sphere packing bounds, error-correcting codes, and the construction of the first explicit non-sofic group — a structure that had remained an open question since 1999. The model also proved the quantum parallel repetition theorem for general two-player entangled games and delivered results on Ehrhart's volume conjecture and Connes' rigidity conjecture. Three additional problems from Paul Erdős's catalog were also resolved, according to reporting by Quartz.qz+2

OpenAI stated that its AI generated the core mathematical arguments, while researchers subsequently formalized each proof using the Lean proof assistant. The GitHub repository's "sorry" count — an indicator of unverified steps — stands at zero, meaning every logical step across all ten proofs is machine-checked.qz

Skeptics Push Back

Despite the formal verification, critics have questioned whether the results represent a true leap in AI capability. Gary Marcus, a prominent AI researcher, wrote that an Anthropic mathematician, Levent Alpöge, was able to partially replicate half of OpenAI's results within 24 hours using Fable, an already publicly released model. Marcus argued this raises doubts about whether Astra itself is the breakthrough or whether the advance lies in identifying which problems are amenable to search-and-verify techniques.garymarcus.substack

"OpenAI does not report which problems they tried and failed at. They have given us a numerator without a denominator, always a worrisome sign," Marcus wrote. He also noted that OpenAI has not yet decided whether to designate Astra as GPT-6 or GPT-5.7, suggesting internally the leap may be considered incremental.garymarcus.substack

The ACM's Communications blog published an analysis on August 4 describing Astra as "amazing — but vastly oversold". Meanwhile, the Leiden Declaration on AI in mathematics, signed by hundreds of mathematicians in June 2026, has flagged broader concerns about unreliable AI-generated proofs and attribution.analyticsinsight+1

What Comes Next

OpenAI has not announced a release date for Astra, and the model will need to clear the federal AI safety review before reaching users. The company acknowledged the debate, stating it has "deep respect and understanding for those concerned with its impact" and plans to publish its methodology. OpenAI also recently launched ChatGPT for Academic Researchers, providing 100,000 scientists free access to its latest models.analyticsinsight+1

Mathematician Terence Tao, in a July 26 lecture referenced by Marcus, drew a distinction between solving isolated open problems — where AI appears strong — and building mathematical theory, where no evidence of AI capability yet exists.garymarcus.substack