Codex CLI is built around OpenAI's GPT-5 models. GPT-5.3-Codex is the coding-tuned default that powers most local runs; GPT-5.5 is the pick for the hardest reasoning; GPT-5.4 is the value sweet spot; and GPT-5.4 Mini handles cheap, high-volume steps.
The coding-tuned default — built for sandboxed execution, diffs, and long agentic coding sessions in Codex.
The flagship for the hardest reasoning and the trickiest refactors, when quality outweighs cost.
The value sweet spot — most of 5.5's ability for under half the price, with a 1M context.
For routing, classification, and the many small calls in a run where a flagship would be wasteful.
The model picks the moves; the agent runs the loop, the tools, and the guardrails. Once you've chosen a model, see which agent gets the most out of it.