OpenAI o3 and o4-mini

2025-04-16 · Publisher : OpenAI · Domain : Artificial Intelligence

OpenAI launches o3, its most powerful reasoning model, and o4-mini, a fast and economical version. For the first time, these models can combine all of ChatGPT's tools and reason about images.

What changes

  • o3 sets new announced records on Codeforces, SWE-bench (without a specific scaffold) and MMMU, with 20% fewer major errors than o1 according to external experts.
  • o4-mini is presented as the best model evaluated on AIME 2024 and 2025, with usage limits significantly higher than o3.
  • Both models decide when to call web search, Python, file analysis and image generation, and “think with images” (zoom, rotation).
  • Launch of Codex CLI, an open-source coding agent in the terminal.
  • o3-pro is announced for a few weeks later (released on June 10, 2025).

OpenAI launches o3, its most powerful reasoning model, and o4-mini, a fast and economical version. For the first time, these models can combine all of ChatGPT's tools and reason about images. ### What changes - o3 sets new announced records on Codeforces, SWE-bench (without a specific scaffold) and MMMU, with 20% fewer major errors than o1 according to external experts. - o4-mini is presented as the best model evaluated on AIME 2024 and 2025, with usage limits significantly higher than o3. - Both models decide when to call web search, Python, file analysis and image generation, and “think with images” (zoom, rotation). - Launch of Codex CLI, an open-source coding agent in the terminal. - o3-pro is announced for a few weeks later (released on June 10, 2025). ### Specifications - Context 200,000 tokens, maximum output 100,000 tokens, text and image input, text output. - Knowledge cutoff: June 1, 2024. - API identifiers: o3 (o3-2025-04-16) and o4-mini (o4-mini-2025-04-16), via Chat Completions and Responses; ChatGPT Plus, Pro and Team from day one, Enterprise and Edu one week later. ### Pricing Launch pricing per million tokens (USD), OpenAI pricing page of May 1, 2025. | Model | Input | Cached input | Output | | --- | --- | --- | --- | | o3 | 10.00 | 2.50 | 40.00 | | o4-mini | 1.10 | 0.275 | 4.40 | The price of o3 was lowered to 2 and 8 dollars on June 10, 2025. ### Benchmarks Scores published by the vendor in the text of the announcement. | Measure | o3 | o4-mini | | --- | --- | --- | | AIME 2025 with Python (pass@1) | 98.4% | 99.5% | | AIME 2025 with Python (consensus@8) | 100% | 100% | The results on GPQA, SWE-bench or MMMU are charts. Taken from the August 2025 GPT-5 announcement (high effort): SWE-bench Verified 69.1% (o3) and 68.1% (o4-mini), GPQA Diamond 83.3% and 81.4%. ### Key takeaways o3 and o4-mini inaugurate agentic reasoning models. o4-mini offers the best cost-performance ratio for high-volume uses.

Links

Tags : LLM, OpenAI, o-series, Raisonnement, Multimodal, Agent