GPT-5.3 Codex vs GPT-5.4

Not a same-generation comparison — GPT-5.3 Codex is the agentic-coding model, GPT-5.4 the general model from the generation after it — and two claims this page used to make were wrong. It said GPT-5.4 had the more recent knowledge cutoff; both are published as the same date. And it called GPT-5.3 Codex the only model with published reasoning-effort control; GPT-5.4's own page lists its levels too.

What actually differs

Context window is the clear one: GPT-5.4's is much larger. Price runs the other way — GPT-5.3 Codex is cheaper on input and cached input, and only slightly cheaper on output. Both sit in the same rate-limit table, so throughput isn't part of the decision.

Effort levels and defaults

GPT-5.4 accepts none through xhigh and defaults to none: an unspecified request does no reasoning. GPT-5.3 Codex accepts low through xhigh and marks no default.

Delivery options

GPT-5.4 publishes Batch and Flex prices and a long-context rule; GPT-5.3 Codex publishes neither — only Standard and Fast, with the long-context rule recorded as not published rather than assumed.

Where they still run

In Codex with ChatGPT sign-in, neither is a current option: GPT-5.4 retired there on August 31, 2026, and GPT-5.3 Codex is listed as deprecated. With your own API key, both remain selectable — GPT-5.3 Codex only until its API shutdown on April 1, 2027, while GPT-5.4 has no announced API shutdown.

How to decide

Choose GPT-5.4 if your work needs its window or Batch pricing; GPT-5.3 Codex only if it fits the smaller window, input cost dominates, and the work will move before its API shutdown. In Codex, compare both against GPT-6.1 Sol first.