First Submission of a Proof for First Proof

OpenAI entered an internal model in the First Proof challenge, which tests whether AI can solve problems in specialized fields and generate verifiable proofs. This challenge goes beyond simple short-answer questions, dealing with advanced, research-level mathematical problems whose correctness is difficult to confirm without expert review.

According to results submitted on February 14, 2026, of the 10 proofs the model attempted, experts assessed that at least 5 (Problems 4, 5, 6, 9, and 10) were likely correct. Problem 2 was initially expected to be correct, but further analysis revealed it to be incorrect.

This experiment was conducted to evaluate the core capabilities of next-generation AI models. It focused on testing advanced reasoning abilities that go beyond simple benchmarks, including:

• Maintaining long chains of reasoning
• Selecting appropriate abstractions
• Handling ambiguity in problem definitions
• Constructing logic that withstands expert review

OpenAI researcher James R. Lee stated that he witnessed the intelligence of the new model—currently in training to enhance the rigor of its thinking—improve in real time during the training process. This achievement is an extension of research that includes reaching IMO (International Mathematical Olympiad) gold medal-level performance and accelerating scientific research using GPT-5.