benchmarks
GPT-5 is OpenAI's reasoning model for coding, analysis, and multimodal understanding, accepting text and images and producing text. The specifications and published launch evaluations below concern the original gpt-5 API model, not other GPT-5-family variants.
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 | 86.2%29 Oct 2025 | View |
| SWE-bench VerifiedResolving real GitHub issues | 73.6%6 Feb 2026 | View |
| OTIS Mock AIME 2024-25Competition mathematics | 91.4%29 Oct 2025 | View |
| Terminal-BenchLong-horizon command-line agent tasks | 49.6%7 Aug 2025 | View |
| FrontierMath | 32.4%13 Nov 2025 | View |
| SimpleQA VerifiedShort factual questions (hallucination resistance) | 50.1%27 Aug 2026 | View |
| MATH Level 5 | 98.1%29 Oct 2025 | View |
| LMArena TextHuman preference Elo from blind chat votes (not a percentage) | 1406 Elo22 Oct 2026 | View |
| Aider PolyglotCode editing across six programming languages | 88%23 Aug 2025 | View |
| LMArena WebDevHuman preference Elo for building web apps (not a percentage) | 1418 Elo23 Oct 2026 | View |
Weighted mean of 8 verified benchmark results: GPQA Diamond 86.2%, SWE-bench Verified 73.6%, OTIS Mock AIME 2024-25 91.4%, Terminal-Bench 49.6%, FrontierMath 32.4%, SimpleQA Verified 50.1%, MATH Level 5 98.1%, Aider Polyglot 88%.
Capabilities
- Software development, repository-level debugging, and code editing.
- Multi-step mathematical and scientific reasoning with configurable reasoning effort.
- Understanding images alongside text, including visual material in analytical tasks.
- Function calling and multi-step workflows involving external tools.
- Long-context document analysis with a documented 400,000-token context window.
- Structured outputs for applications requiring schema-constrained responses.
What improved
- OpenAI reports fewer factual errors than GPT-4o and o3 under its evaluated conditions.
- Published coding evaluations show stronger software-engineering and code-editing performance than o3.
- OpenAI reports improved instruction following and multi-step tool use.
- Training targets reduced sycophancy and more accurate acknowledgment of limitations.
- Safe-completion training aims to preserve useful assistance while withholding unsafe details.
Known flaws & limitations
- Can still generate incorrect facts, fabricated references, and faulty technical explanations.
- Prompt-injection resistance remains incomplete when handling untrusted content.
- Benchmark results depend on reasoning effort, tools, evaluation scaffolding, and task conditions.
- The gpt-5 API model does not natively support audio or video input or output.
- Strong benchmark results do not establish reliable autonomous performance on unfamiliar, long-running tasks.
Risks, safety & security issues
- Malicious instructions embedded in documents or tool responses can attempt to redirect application behavior or expose data.
- Unreviewed errors can cause harm in medical, legal, financial, and other consequential decisions.
- Coding and automation capabilities can be misused for malicious software development or social engineering.
- OpenAI applied heightened biological and chemical safeguards under its Preparedness Framework.
- Broad tool permissions can amplify damage from mistaken or manipulated actions.
Good fit for
- Human-reviewed software development, debugging, and code transformation.
- Document synthesis and image-assisted analysis where important claims can be checked.
- Mathematical and scientific problem-solving with independent validation.
- Tool-assisted workflows with restricted permissions and approval checkpoints.
Poor fit for
- Unsupervised high-stakes decisions requiring guaranteed factual correctness.
- Autonomous execution of irreversible actions with unrestricted credentials.
- Applications requiring native audio or video processing.
- Workflows that treat instructions inside untrusted documents as authoritative.
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.