benchmarks
OpenAI o3 is a reasoning model supporting text and image inputs, text outputs, and tool-assisted problem solving. Its documented strengths include scientific reasoning, coding, and visual analysis; o3-pro is a separate, higher-compute variant released after o3.
Benchmark results
Every figure links to where it was published. Compare like with like — harnesses and prompt setups differ between vendors.
| Benchmark | Score | Source |
|---|---|---|
| GPQA DiamondPhD-level science questions | 81.8%16 Apr 2025 | View |
| SWE-bench VerifiedResolving real GitHub issues | 62.3%12 Feb 2026 | View |
| OTIS Mock AIME 2024-25Competition mathematics | 84.4%7 Aug 2026 | View |
| FrontierMath | 18.7%16 Nov 2025 | View |
| SimpleQA VerifiedShort factual questions (hallucination resistance) | 49.4%27 Aug 2026 | View |
| MATH Level 5 | 97.8%16 Apr 2025 | View |
| LMArena TextHuman preference Elo from blind chat votes (not a percentage) | 1410 Elo22 Oct 2026 | View |
| Aider PolyglotCode editing across six programming languages | 81.3%25 Jun 2025 | View |
| LiveCodeBenchFresh competitive-programming problems | 84.7%1 May 2025 | View |
Weighted mean of 8 verified benchmark results: GPQA Diamond 81.8%, SWE-bench Verified 62.3%, OTIS Mock AIME 2024-25 84.4%, FrontierMath 18.7%, SimpleQA Verified 49.4%, MATH Level 5 97.8%, Aider Polyglot 81.3%, LiveCodeBench 84.7%.
Capabilities
- Multi-step mathematical and scientific reasoning.
- Code generation, debugging, and software issue resolution.
- Visual reasoning over images, charts, and diagrams.
- Tool-assisted workflows combining reasoning with external information and computation.
- Function calling and structured outputs for application integration.
What improved
- Higher GPQA Diamond performance than o1 in OpenAI's published evaluation.
- Higher SWE-bench Verified performance than o1 under OpenAI's reported evaluation setup.
- Higher MMMU performance than o1 in OpenAI's published visual-reasoning evaluation.
- OpenAI reported 20% fewer major errors than o1 in expert evaluations of difficult real-world tasks.
Known flaws & limitations
- Can produce confident factual errors; OpenAI reported higher hallucination rates than o1 on some factuality evaluations.
- Extended reasoning can increase response latency and consume substantial output tokens.
- Its documented knowledge cutoff is June 2024, so newer facts require external information.
- Supports text and image inputs, but not native audio or video inputs.
Risks, safety & security issues
- Confident factual errors can mislead users in consequential decisions without independent validation.
- Untrusted documents and tool results can introduce prompt-injection risks in connected applications.
- Advanced coding and scientific capabilities create dual-use risks despite safety training and safeguards.
- Connecting sensitive data or privileged tools can increase the consequences of incorrect or manipulated actions.
Good fit for
- Complex scientific and mathematical analysis with independently checked conclusions.
- Software debugging and repository-level issue resolution with automated tests and code review.
- Interpreting charts, diagrams, and other visual technical material.
- Multi-step research and analysis using controlled tools and cited external sources.
Poor fit for
- High-stakes factual decisions without expert review or source verification.
- Strictly latency-sensitive tasks that do not require extended reasoning.
- Native real-time audio or video processing.
- Unsupervised actions involving sensitive data, payments, or privileged systems.
References
Profile last reviewed 10/27/2026 and re-checked automatically every few hours against the vendor's published documentation, papers and public leaderboards. Scores come from the cited benchmark results above — confirm limits and pricing against the provider's own documentation before building on them.