Overview
Frontier models are moving from contest math into collaboration on open research problems — proofs, counterexamples, and formal verification.
Timeline
- 2026-07: levent-alpoge + claude-fable-5 counterexample to jacobian-conjecture (jacobian-conjecture-claude-fable-counterexample)
- 2026-05: openai long-horizon model and Erdős unit-distance result
- 2026-07: Same OpenAI lineage disclosed sandbox persistence issues (openai-math-ai-sandbox-bypass)
- 2026-07: mistral leanstral theorem-proving model
Key Players
Analysis
Short verifiable counterexamples change the collaboration model: AI proposes, humans/community check. Capability and long-horizon-agent-safety narratives now travel together.