GPT-5-Codex
OpenAI's GPT-5 variant tuned for autonomous coding agents — 7+ hour task horizons with adaptive reasoning.
About
GPT-5-Codex (September 15, 2025) is a GPT-5 variant purpose-built for agentic software engineering, trained with reinforcement learning on real-world tasks: building projects from scratch, adding features and tests, debugging, and large-scale refactoring. Its signature behavior is adaptive reasoning — spending 93.7% fewer tokens than GPT-5 on the simplest 10% of requests and roughly twice the reasoning effort on the hardest 10%. OpenAI reported it working autonomously for over seven hours and refactoring a pull request across 232 files and 3,500 lines in the Gitea repo (provider-reported). Code review got dedicated attention: repo navigation, dependency analysis, and test runs, with incorrect review comments cut from 13.7% to 4.4% versus GPT-5 (provider-reported). It became the default across the Codex CLI, IDE extension, cloud service, and GitHub code review integration, plus the ChatGPT app for subscribers; API access was promised but not yet shipped at launch.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- SWE-bench 74.5% — provider-reported via press
- Refactoring accuracy 51.3% vs GPT-5 33.9% — provider-reported
- Incorrect code-review comments 13.7% to 4.4% vs GPT-5 — provider-reported
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Sep 15, 2025 launch date consistent across OneYearAgo, Blockchain.News, Implicator, CIOL; Adaptive reasoning figures (93.7% fewer tokens, 7+ hour autonomy, 232-file refactor) provider-reported via press — labeled as such; SWE-bench 74.5% via secondary press citing OpenAI — labeled provider-reported.