The cheaper sibling of GPT-5.4 — same 272K window, an eighth of the output price, built for well-defined tasks.
GPT-5.4 mini is the middle child of the GPT-5.4 family: smarter than nano, an eighth of the output price of full GPT-5.4, with the same 272K context window. OpenAI also uses it as the default model on ChatGPT's free tier, which tells you the positioning — good enough for everyday tasks at a cost OpenAI can give away.
On the API, the pitch is "well-defined tasks": summarization, extraction, classification, formulaic generation. When the prompt fully specifies the job, mini does it at 8–10× less than the full model. When the task needs planning or multi-step reasoning, the savings evaporate in retries — that's full GPT-5.4 territory.
Mini handles the same multimodal surface as its bigger sibling (vision in, structured output, tool calling) and is fast enough for interactive products. The honest weakness: it follows instructions more literally and plans less. Agent loops longer than a couple of steps, ambiguous prompts, and subtle code work all favor the full model.
| Model | Input / 1M | Output / 1M | Context |
|---|---|---|---|
| GPT-5.4 mini | $0.25 | $2 | 272K |
| GPT-5.4 | $2.50 | $15 | 272K |
| GPT-5.4 nano | $0.05 | $0.40 | 272K |
| Gemini 3.1 Flash-Lite | $0.25 | $1.50 | 1M |
| Claude Haiku 4.5 | $1 | $5 | 200K |
| Mistral Small 3.1 | $0.20 | $0.60 | 128K |
Cross-family, Gemini 3.1 Flash-Lite matches mini's input price with cheaper output and a 1M window; Mistral Small undercuts it on both rates. Mini's draw is ecosystem: same API shape, tooling, and behavior as the rest of the GPT stack, so a GPT-5.5 product can route easy traffic to mini with one parameter change.