GPT-5.4 — Specs, Pricing & Limits

GPT-5.4 now sits behind both the GPT-5.6 and GPT-6 generations, and the most practical fact about it is where it still works. It left Codex for ChatGPT sign-in on August 31, 2026 — OpenAI's Codex models page names GPT-6 Sol as the replacement there — but it remains on the API, and Codex run with your own API key can still select it.

Its context window and rate-limit group match the GPT-6 Sol and Astra rows exactly, and its long-context rule is worded "for the full session for standard, batch, and flex": crossing the threshold reprices the whole request, and there is no long-context row for its Fast tier. Its knowledge cutoff sits well behind every GPT-5.6 and GPT-6 model — recency, more often than price, is what pushes a migration off this model.

Two details catch people auditing a bill. First, it is the only frontier-sized model whose published default reasoning effort is none: an unspecified request gets no reasoning at all, unlike the GPT-5.6 and GPT-6 models, which default to medium wherever their pages mark a default. Second, its Batch cached-input rate is published as a rounded figure that halving the Standard rate won't reproduce exactly — this site ships the published cell, so a script computing it independently will disagree by a fraction of a cent.

It adds no cache-write surcharge: a token written to the cache bills as ordinary input, the rule for every model before GPT-5.6.

The table above is computed directly from this site's facts module: pricing across every published service tier, the long-context bands, and the rate-limit table.