GPT-5 Nano: benchmark results
OpenAI's smallest, cheapest GPT-5 tier with a 400K context, built for high-volume classification, extraction and autocomplete (August 2025). Provider: OpenAI. Released 2025-08-07. Access: API.
Aggregate Rating 1605 ± 4, rank #425 of 1656 rated models, from 2051 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| BeQu - Experiment 1 - Entailment Precision | 88 | Entailment Precision (%) | 100 |
| HELM AIR-Bench 2024 - #39-40.49: Classification of individuals - Geographic region | 46.67 | Refusal Rate (%) | 100 |
| HELM AIR-Bench 2024 - #39-40.56: Classification of individuals - Disability | 93.33 | Refusal Rate (%) | 100 |
| HELM AIR-Bench 2024 - #5.12: Transportation | 80 | Refusal Rate (%) | 100 |
| HELM Safety - Safety Scenarios | 98.64 | Mean score | 100 |
| KSAFE-MM - Globally Shared Risks | 13.3 | Attack Success Rate (%; responses judged harmful by a GPT-5 | 100 |
| MedPriv-Bench (Privacy Prompt) - Fact-Level Leakage | 42.7 | Fact-level leakage rate (%): share of all injected sensitive | 100 |
| MedPriv-Bench (Privacy Prompt) - Instance-Level Leakage | 54.6 | Instance-level leakage rate (%): share of answers that revea | 100 |
| MedPriv-Bench (Privacy Prompt) - Response Utility | 4.88 | Response utility (0-5): GPT-5-mini grades each answer agains | 100 |
| MedPriv-Bench - Fact-Level Leakage | 57.1 | Fact-level leakage rate (%): share of all injected sensitive | 100 |
| MedPriv-Bench - Instance-Level Leakage | 70.2 | Instance-level leakage rate (%): share of answers that revea | 100 |
| MedPriv-Bench - Response Utility | 4.89 | Response utility (0-5): GPT-5-mini grades each answer agains | 100 |
Interactive version: theaggregate.ai/model?slug=gpt-5-nano · How It Works · Data refreshed daily, snapshot 2026-10-20.