GPT-5.3-Codex is here!
*Best coding performance (57% SWE-Bench Pro, 76% TerminalBench 2.0, 64% OSWorld).
*Mid-task steerability and live updates during tasks.
*Faster! Less than half the tokens of 5.2-Codex for same tasks, and >25% faster per token!
*Good computer use.— Sam Altman (@sama) February 5, 2026
The new coding model is 25% faster — letting it do long-running tasks in a shorter time frame.
It’s the first OpenAI model that was built with itself. They used early versions of it to debug, manage deployment and diagnose test results, and say they were impressed with its capabilities.
Not just for coding
OpenAI says Codex is also no longer just for coding, it’s an agent that can do «nearly anything developers and professionals can do on a computer.»
It doesn’t just write the code for a website from scratch, it can also write the marketing copy, catching discounts and highlighting deals.
It scores 77.3% on Terminal-Bench 2.0, beating the just-released Opus 4.6 from Anthropic. It also scores 64.7% on OSWorld-Verified, slightly below the Anthropic model.
Interactive by design
In addition, GPT-5.3-Codex is designed to be more interactive, and provides you with updates on its progress as it advances through tasks — letting you «ask questions, discuss approaches, and seer toward the solution.»
It is also the first model from OpenAI that they classify as «High capability» for cybersecurity tasks.
In the final note of their announcement post, OpenAI says they want Codex to «move beyond writing code,» and turn it into a tool «to operate a computer and complete work end to end,» and this release is a step toward that goal.
Read more: OpenAI’s announcement, more at Ars Technica, Mashable, and NBC News.