OpenAI unveils GPT-5-Codex, a coding-tuned variant of GPT-5

OpenAI has introduced GPT-5-Codex, a version of GPT-5 further optimized for agentic coding in Codex, according to an addendum to the GPT-5 system card. The disclosure extends the safety and capability documentation the company published alongside GPT-5 itself.

The defining change is how the model allocates its own compute. GPT-5-Codex adjusts its thinking effort more dynamically based on task complexity, the addendum states. It responds quickly to simple conversational queries or small tasks, while working independently for longer stretches on more complex ones.

That behavior matters for developers. Agentic coding tools increasingly run unsupervised for minutes or hours at a time, and a model that spends disproportionate reasoning on trivial requests wastes time and money, while one that underthinks hard problems ships broken code. Adaptive effort allocation is one way vendors are trying to balance speed, cost, and reliability in autonomous workflows.

The name anchors the model to a specific product context. Codex is OpenAI's agentic coding environment, and GPT-5-Codex is tuned for that setting rather than positioned as a general-purpose release. OpenAI frames it as "a version of GPT-5 further optimized for agentic coding in Codex" — language that signals an incremental specialization of the flagship model rather than a new generation.

The choice of a system card addendum as the release vehicle also carries weight. System cards are OpenAI's primary mechanism for documenting model behavior, capabilities, and risks ahead of or alongside deployment. By documenting GPT-5-Codex in an addendum to the existing GPT-5 card rather than in a standalone document, the company treats the new model as a derivative of an already-assessed system, a decision that will likely draw attention from policy watchers tracking how AI developers scope and disclose model variations.

The coding-agent market is currently one of the most competitive segments in AI, with developers judging tools on latency for simple requests and stamina for long-running, multi-step engineering tasks. GPT-5-Codex's dynamic effort mechanic addresses both ends of that spectrum in a single model. Whether the approach delivers measurable gains over fixed-effort configurations is something developers will test once they get hands-on time with the model in Codex.