GPT-6 Astra Solved FrontierMath Tier4's Last Unsolved Problem; Benchmark Now Saturated
量子位 · wechat · 2026-09-11
GPT-6 Astra cracked the only Tier4 problem never solved by AI, and EpochAI has declared FrontierMath Tier4 saturated, with OpenAI reporting 97.6% versus 90.2% for Claude Fable 5 and 83.0% for GPT-5.6 Sol.
- FrontierMath launched in Nov 2024 with 60+ mathematicians (including Tao and Gowers) specifically to resist benchmark saturation; top models scored under 2% at first
- Tier4 (research-level problems from professors' own work) debuted in July 2025 with only 3 problems ever solved; Epoch once wrote some might take decades
- After OpenAI flagged errors, Epoch audited and shipped v2 in June 2026, fixing 12 and removing 7 problems, leaving 43
- Problem author Jay Pantone says Astra's solution was close to a human mathematician's, not a numerical shortcut
- The project is moving on: OpenProblems and FrontierMath-Erdős (Lean-formalized proofs), where Astra solved just 2 of 68
More from Models
- Model refuses wrong psychiatric advice in deliberately misleading test — FellMentKE · 2026-10-27
- Clinical System One v0.2: top clinical AI drops from 97% to 65% on harder cases — MaziyarPanahi · 2026-10-27
- ChatGPT keeps telling people 404 Media is tech support for big corporations — 404 Media · 2026-10-26
- Claude's /show-me visualizes compressor auto makeup gain like a pro audio teacher — teropa · 2026-10-26
- ChatGPT keeps telling people 404 Media can fix their tech support problems — 404 Media · 2026-10-26
- Fern's "context is intelligence" thesis: prediction keeps improving out to 10.2M tokens — bneyshabur · 2026-10-26