Now, OpenAI is launching a custom, optimized version of GPT-5 for the Codex coding agent that they say is faster, more reliable and more steerable than before.
Slightly better than vanilla
The new model scores slightly better than GPT-5 (high) on SWE-bench Verified at 74.5% versus 72.8%.
GPT-5 Codex also thinks a lot longer on tough tasks, and was seen working independently for over 7 hours at a time in testing.
At the same time, OpenAI is upgrading the CLI, the IDE and making changes to the infrastructure to make it faster and more efficient.
Also of note is that this model isn’t available in the API just yet, with OpenAI saying «For developers using Codex CLI via API key, we plan to make GPT‑5-Codex available in the API soon.»
Read much more at OpenAIs announcement, and see also the story at TechCrunch. There is also an OpenAI podcast discussing GPT-5 Codex.