OpenAI released o1-preview, a model trained via large-scale reinforcement learning to improve chain-of-thought reasoning. On the 2024 AIME exam, o1 achieved 74% single-sample accuracy versus GPT-4o's 12%; on GPQA Diamond (PhD-level science), o1 surpassed recruited human PhD experts; and on Codeforces competitive coding, o1 ranked in the 89th percentile. Performance scales smoothly with both train-time compute (reinforcement learning steps) and test-time compute (longer thinking). This demonstrates measurable improvement in multi-step reasoning and error correction beyond what standard scaling alone has achieved. For financial institutions, this signals material progress in agentic reasoning capability that could affect enterprise ROI assumptions and the timeline for knowledge-worker displacement in reasoning-heavy occupations.
245 other tracked developments touching the same claims.