OpenAI o3

Released April 16, 2025 alongside o4-mini, o3 was OpenAI's most powerful reasoning model of its generation, built on large-scale reinforcement learning over chains of thought. Its defining trait is agentic reasoning: it interleaves web browsing, Python execution, image manipulation, and file analysis inside a single extended reasoning trace rather than treating tools as separate steps. It set new state of the art on SWE-bench Verified (~69-72%), Codeforces (2727 Elo), and GPQA Diamond (~88%), and was the first OpenAI reasoning model to natively reason with images.\n\nLimitations: it hallucinated more than o1 — OpenAI's own tests showed higher claim-generation error rates on person-knowledge benchmarks. Reasoning traces can take minutes and cost significantly more than non-reasoning models. OpenAI has since folded the o-line into GPT-5-era models, so o3 is a legacy checkpoint rather than a maintained line.

OpenAI-reported: ~69.1% SWE-bench Verified, 2727 Codeforces Elo, 87.7% GPQA Diamond