GPT-5
GPT-5 unifies fast and reasoning models behind a real-time router in ChatGPT and ships as gpt-5, mini and nano in the API.
In ChatGPT a router picks between a fast model and GPT-5 thinking; in the API the three sizes take reasoning effort from minimal to high, with a 400K window. OpenAI reports 94.6% on AIME 2025 and 74.9% on SWE-bench Verified, using 50-80% fewer output tokens than o3. Priced $1.25/$10 per million tokens.
- Date
- Thursday, 7 August 2025
- Lab
- OpenAI
- Kind
- model
- Access
- closed API
- Price
- gpt-5 $1.25/$10 per 1M tokens (2025-08-07)
Figures
| Measure | Value | Measured by |
|---|---|---|
| SWE-bench Verified | 74.9% company-reported; Aider polyglot 88% | company |
| AIME 2025, no tools | 94.6% | company |
| MMMU | 84.2% | company |
| Factual-error reduction with search | ~45% fewer vs GPT-4o; ~80% fewer vs o3 when thinking company-run on ChatGPT-like production prompts | company |
| API price GPT-5 / mini / nano (input per 1M tokens) | $1.25 / $0.25 / $0.05 output $10 / $2 / $0.40 | company |
Launch was rocky. Removal of GPT-4o from ChatGPT triggered a backlash and a same-week reversal (Simon Willison, 2025-08-08), and router glitches were acknowledged. Treated as High capability in biology. GPT-4o was finally retired from ChatGPT on 2026-02-13.
Sources
- openai.com/index/introducing-gpt-5/
- openai.com/index/introducing-gpt-5-for-developers/
- simonwillison.net/2025/Aug/7/gpt-5/
- deploymentsafety.openai.com/gpt-5/
This record was checked against its sources on 6 October 2026. How we check