GPT-5.3-Codex
GPT-5.3-Codex is OpenAI's first model "instrumental in creating itself", as early versions helped debug its own training and manage its deployment.
Merges GPT-5.2-Codex coding with GPT-5.2 knowledge work in one model that is 25% faster. Terminal-Bench 2.0 rises from 64.0% to 77.3%; OSWorld-Verified from 38.2% to 64.7%. First model OpenAI classifies High for cybersecurity, with $10M in API credits for defenders. API followed 2026-02-24.
- Date
- Thursday, 5 February 2026
- Lab
- OpenAI
- Kind
- model
- Access
- closed API
- Price
- API from 2026-02-24
Figures
| Measure | Value | Measured by |
|---|---|---|
| Terminal-Bench 2.0 | 77.3% GPT-5.2-Codex 64.0%, GPT-5.2 62.2% | company |
| SWE-Bench Pro (public) | 56.8% GPT-5.2-Codex 56.4% | company |
| OSWorld-Verified | 64.7% GPT-5.2-Codex 38.2% | company |
Released within about 15 minutes of Anthropic's Opus 4.6 (Simon Willison). The self-development claim is OpenAI's own description of internal use, not independently audited. Also launched Trusted Access for Cyber.
Sources
- openai.com/index/introducing-gpt-5-3-codex/
- deploymentsafety.openai.com/gpt-5-3-codex/
- simonwillison.net/2026/Feb/5/two-new-models/
- developers.openai.com/api/docs/changelog
This record was checked against its sources on 6 October 2026. How we check