benchmarks
GPT-5 nano is a compact GPT-5-family reasoning model designed for low-cost, low-latency tasks, particularly summarization and classification. It accepts text and image inputs and produces text outputs, with configurable reasoning effort and a documented 400,000-token context window.
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 | 69.4%30 Oct 2025 | View |
| OTIS Mock AIME 2024-25Competition mathematics | 81.1%31 Oct 2025 | View |
| Terminal-BenchLong-horizon command-line agent tasks | 21.8%7 Aug 2025 | View |
| FrontierMath | 8.3%30 Oct 2025 | View |
| SimpleQA VerifiedShort factual questions (hallucination resistance) | 11.7%10 Aug 2026 | View |
| MATH Level 5 | 95.2%20 Aug 2025 | View |
Weighted mean of 6 verified benchmark results: GPQA Diamond 69.4%, OTIS Mock AIME 2024-25 81.1%, Terminal-Bench 21.8%, FrontierMath 8.3%, SimpleQA Verified 11.7%, MATH Level 5 95.2%.
Capabilities
- Low-cost summarization and classification for high-volume text workflows.
- Text and image input with text output.
- Configurable reasoning effort for balancing response speed and reasoning depth.
- Function calling and structured outputs for application integration.
- A 400,000-token context window and up to 128,000 output tokens.
What improved
- Standard input pricing is 50% lower than GPT-4.1 nano: $0.05 instead of $0.10 per million tokens.
- Cached-input pricing is 80% lower than GPT-4.1 nano: $0.005 instead of $0.025 per million tokens.
- Adds configurable reasoning effort compared with the non-reasoning GPT-4.1 nano.
- Raises the documented maximum output from GPT-4.1 nano's 32,768 tokens to 128,000 tokens.
Known flaws & limitations
- The documented knowledge cutoff is May 31, 2024; newer facts require external information.
- Native audio and video inputs are not supported.
- Output is text-only; native image and audio generation are not supported.
- The 400,000-token context window is smaller than GPT-4.1 nano's 1,047,576-token window.
- OpenAI's launch evaluations show lower GPQA Diamond and AIME 2025 performance than GPT-5 mini and GPT-5.
Risks, safety & security issues
- Prompt injection remains a GPT-5-family concern when applications process untrusted content or expose tools.
- Incorrect or fabricated answers can mislead users when generated claims are treated as verified facts.
- Model-level safety protections do not eliminate harmful-output risks or replace application-level access controls.
Good fit for
- High-volume text classification, tagging and routing.
- Summarization of documents that fit within the context window.
- Structured information extraction from text and images.
- Cost-sensitive tool-assisted workflows with constrained permissions and output validation.
Poor fit for
- Native audio or video processing and image or audio generation.
- High-stakes decisions without independent verification and qualified human review.
- Tasks requiring current factual knowledge without retrieval or other external information.
- Workloads requiring approximately one million tokens in a single context.
References
Profile last reviewed 10/18/2026 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.