o3
By OpenAI
o3 is OpenAI's advanced reasoning model, released in 2025 as the successor to o1, delivering stronger performance on complex math, coding, and scientific reasoning through more extensive step-by-step internal reasoning.
Definition
o3 is OpenAI's advanced reasoning model, released in 2025 as the successor to o1, delivering stronger performance on complex math, coding, and scientific reasoning through more extensive step-by-step internal reasoning.
Overview
o3 continued OpenAI's reasoning-model line, building on o1 with improvements in accuracy on some of the hardest available reasoning and coding benchmarks at the time of release, including strong results on competition mathematics and PhD-level science questions. OpenAI also released a lighter, faster sibling, o3-mini, offering a similar reasoning approach at lower cost and latency for less demanding tasks. Like o1, o3 allocates variable amounts of internal 'thinking' computation depending on task difficulty, and this configurable reasoning effort lets developers trade off cost and latency against accuracy for a given application. o3 also improved tool use during reasoning, allowing the model to more effectively interleave steps like running code or searching before finalizing an answer, which strengthened its usefulness for AI agent style workflows. The reasoning techniques pioneered in o1 and o3 were subsequently integrated into OpenAI's unified GPT-5 model, which automatically decides when to apply o3-style deep reasoning versus a fast response, rather than requiring users to pick a separate reasoning model.
Key Concepts
- Successor to o1 with improved reasoning accuracy
- Strong results on advanced math, coding, and science benchmarks
- Lighter o3-mini variant for lower-cost, lower-latency reasoning
- Configurable reasoning effort for cost/accuracy tradeoffs
- Improved integration of tool use (e.g., code execution) during reasoning
- Reasoning approach later folded into GPT-5's unified routing