OpenAI o3-pro

o3-pro, released June 2025, is not a new architecture — it's the o3 reasoning model given substantially more inference compute: longer internal reasoning chains and more thorough self-checking before answering. It scored 93% on AIME 2024 and 84% on GPQA Diamond at launch, beating standard o3 and topping Gemini 2.5 Pro and Claude 4 Opus on those vendor-reported benchmarks. It's aimed at the slice of tasks where o3 is close but occasionally wrong: hard math proofs, formal reasoning, tough software engineering, and scientific analysis.

The honest picture: pricing is steep ($20/$80 per million tokens), ten times o3's cost, so it's wasteful on anything routine — the extra compute rarely changes easy answers. It's Responses-API-only, which forced codebase migrations for Chat Completions users. Note that OpenAI scheduled the o3-pro-2025-06-10 snapshot for shutdown in December 2026, so check current API availability before building on it.