OpenAI's Astra Cracks Progress on Ten Open Math Problems

a16z released an in-depth conversation with OpenAI mathematicians Mark Sellke and Mehtaab Sawhney, along with a16z infrastructure partner Lisha Li, revealing a series of results achieved by OpenAI's math research system Astra around the GPT-6 Astra launch: breakthroughs emerged one after another on roughly ten open problems, including Navier–Stokes equation solutions, progress on the twin prime conjecture, sphere packing (refreshing a 1970s-era upper bound), and the existence of non-sofic groups. Multiple posters described it as an extraordinary stretch of weeks for AI mathematics.

Confirmed

• The conversation was recorded before GPT-6 Astra's release; Sawhney had fed the system an Erdős problem that was still listed as open at the time as a starting point, and the related results were subsequently made public
• Posters relaying the discussion said the reasoning trajectories closely resemble how human experts do mathematics
• The results also include constructing non-sofic groups and sphere packing progress—previously unresolved problems

Why It Matters

• Multiple long-standing open problems were advanced by an AI system in a short time, seen as a landmark in AI mathematical capability and one of the core highlights of the Astra launch
• The conversation offers a firsthand perspective from OpenAI researchers on how AI can participate in original mathematical research

2026-09-08 ~ 2026-09-08 · 5 related posts

• [source] OpenAI Mathematicians: Model Reasoning Traces Look Strikingly Like Expert Humans' — a16z · 2026-09-08
• Astra's math blitz: Navier-Stokes, twin primes and ten open problems, explained by OpenAI — lishali88 · 2026-09-08
• OpenAI researchers on GPT-6 Astra: original math before launch, 1970s bound improved — a16z · 2026-09-08
• OpenAI's GPT-6 Astra cracked 10 open math problems and improved a 1970s sphere-packing bound — morqon · 2026-09-08

1 near-duplicate retellings: lishali88