GPT 5.4 Mini: the budget OpenAI id and its ceilings

What it costs and what it keeps

gpt-5.4-mini lists at $0.75 per million input tokens and $4.50 output, billed at the 0.1× multiplier — $0.075 and $0.45 in practice. Against gpt-5.4 in the same generation, that is 3.33× cheaper on both sides; against gpt-5.5 it is 6.67× cheaper.

The word "mini" usually implies a stripped feature set. Here it does not: the catalog records text, tool calling and thinking on gpt-5.4-mini, the same capability list as gpt-5.4 and gpt-5.5, and the same /v1/responses route with the same Authorization: Bearer header.

It also shares its generation’s release date, 2026-03-15. So the trade is not "fewer features for less money" — it is the same feature list at a lower rate, with the ceilings described below applying to the whole GPT 5.4/5.5 generation rather than to this id alone.

curl https://token-share.app/v1/responses \
  -H "Authorization: Bearer $TOKEN_SHARE_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.4-mini",
    "input": "Return only one word, bug or feature: the commit message is \"allow empty tags in the parser\"."
  }'

The ceilings: 272k in, 64k out

gpt-5.4-mini takes a 272,000-token context and emits at most 64,000 tokens. Both numbers are the smallest in the OpenAI text lineup here — the GPT 5.6 ids take 372,000 and emit up to 128,000, and stable-gpt-6-astra takes 1,050,000.

The output ceiling is the one that fails loudly. A task that must return a single response longer than 64,000 tokens — a full file rewrite, a long generated document — does not get a cheaper answer here, it gets a truncated one. That is a hard limit, not a quality question.

The context ceiling fails more quietly. 272,000 tokens is a lot of room until a long agent session accumulates tool results in the transcript, at which point the window fills and the cost of dropping to a cheaper id turns into the cost of managing a smaller window.

A sibling that undercuts it

Before settling on gpt-5.4-mini as the budget option, note that gpt-5.6-luna is cheaper per token: $0.06 / $0.36 after its multiplier, against $0.075 / $0.45. Luna also carries the larger 372,000-token context and the higher 128,000-token output ceiling.

The one thing that trips people up is the multiplier column, which points the other way — Luna is 0.3× and gpt-5.4-mini is 0.1×. The multiplier is a discount applied to each model’s own list rate, and Luna starts from a much lower one, so the post-discount price ends up lower despite the larger multiplier.

That does not make the choice automatic. The two are different generations, and the catalog records nothing about relative output quality, so the way to settle it is to run a representative sample of your own prompts through both ids and compare outputs and retry counts. Both take the same request shape on /v1/responses, so that comparison costs one extra call per prompt.

Common questions