OpenAI o3 and o4-mini Review — Reasoning Models That Think With Images

Released April 16, 2025, o3 and o4-mini were OpenAI's most capable models at launch — the first reasoning models to natively use tools, incorporate images into chain-of-thought, and match human performance on graduate-level science. o3 scored 87.5% on ARC-AGI; o4-mini achieved 99.5% on AIME 2025 with code execution. We review architecture, benchmarks, pricing, the FrontierMath controversy, agentic design, and how these models look from 2026.