FrontierMath Erdős: internal OpenAI model solves 16/68 problems, GPT-6.2 Astra just 7/68
thomasahle · x · 2026-11-18
Per Dmitry Rybin, on the FrontierMath Erdős benchmark an internal OpenAI model (unspecified budget) solved 16 of 68 problems, while GPT-6.2 Astra managed only 7 of 68 with a $240-per-problem budget.
FrontierMath Erdős, released by Epoch AI in August 2026, collects 68 open problems posed or studied by Paul Erdős. As of late October 2026, 17 have been solved — with or without AI. The official protocol allows one attempt per problem with a $300 and 72-hour budget; OpenAI released 700+ Lean proofs in October 2026, knocking out many at once. Every evaluated model's knowledge cutoff (June 2026) predates any published solution.
More from Models
- HF adapts Qwen for pi coding agent with RL+LoRA: SWE-bench Pro Hard jumps 29.4%→62.7% — art_zucker · 2026-11-18
- Why AI Is Sycophantic: A Fully Honest AI Would Be Bad for Business — JFPuget · 2026-11-18
- Dev predicts Opus 5.5 will be remembered alongside Sonnet 3.5 and GPT-4 — inductionheads · 2026-11-18
- Ming-3.0-audio-flash Tops Audio Reasoning Benchmark at 99.8% Accuracy — alifcoder · 2026-11-18
- Community Note clash over AI-text detection: 'complete poem shows 100% human' — JFPuget · 2026-11-18
- Switching from ChatGPT Pro to Claude Ultra changed this user's weekly work rhythm — minhnguyen · 2026-11-18