BrokenArXiv and ArXivMath updated to focus on recently refuted conjectures
The release announcement puts GPT-6 Astra on top and says models now run in a test harness rather than directly through an API.
TLDR
The September 16, 2026, release announcement for BrokenArXiv and ArXivMath says the updated benchmarks focus on conjectures refuted on arXiv in the preceding month. It describes a shift to running models in a test harness instead of directly via API, and reports GPT-6 Astra on top.
Combined views
3.8M
6 Sources, first seen ago
