Introducing GPT-5.3-Codex

model release · source date 2026-02-05 · added 2026-08-07 15:52:43 · updated 2026-08-07 15:52:43 · Open original blog

Long-horizon agent work needs both strong coding and strong reasoning, which had been split across separate models.
Fire-and-forget agent runs waste time when the user can see it going wrong but cannot intervene.
Prompting guidance for agent models is under-specified relative to how much it changes behavior.

Combines GPT-5.2-Codex coding with GPT-5.2 reasoning, roughly 25% faster, aimed at long-horizon multi-step work.
Adds real-time interactive steering so a run can be corrected mid-flight rather than restarted.
Positioned beyond code, toward "nearly anything developers and professionals do on a computer."
Paired with an updated prompting guide introducing a required `phase` parameter, and the Codex App Server architecture.

This is the model layer that the harness-engineering, self-improving-tax-agent, and Codex customization work in this scan is built around.
The `phase` parameter is the notable detail: a single prompting knob reported to cause major degradation when omitted, which is a reminder that model-specific harness contracts are real and undocumented ones are a liability.
Reportedly the first model "instrumental in creating itself," used to debug its own training and deployment.
Caveat: a vendor launch. Capability and speed claims, and the self-creation framing, are OpenAI-reported and not independently benchmarked; the `phase` guidance is also first-party.

Comments

No comments yet.