GPT-5.4 Mini Guide: Most of GPT-5.4 at Under a Third the Price

Small models used to be what you reached for when quality did not matter much. GPT-5.4 mini breaks that habit. On the two benchmarks OpenAI chose to headline, it sits a few points behind the full GPT-5.4 at less than a third of the price, and for a large share of agent work that makes it the better default rather than the budget option.

OpenAI released GPT-5.4 mini and GPT-5.4 nano together on 17 March 2026, twelve days after GPT-5.4 itself. The current snapshot is gpt-5.4-mini-2026-03-17.

How close mini gets to the full model

OpenAI's launch figures, as reported at release:

• SWE-Bench Pro: GPT-5.4 at 57.7%, mini at 54.4%, nano at 52.4%, and the older GPT-5 mini at 45.7%
• OSWorld-Verified: GPT-5.4 at 75.0%, mini at 72.1%, nano at 39.0%

Two things stand out. Against the previous small model, GPT-5 mini, the jump on SWE-Bench Pro is close to nine points, so this is not a minor refresh. And on OSWorld-Verified, which tests operating desktop software through screenshots, mini lands near the 72.4% human baseline that the benchmark cites, a place nano does not come close to.

OpenAI also says mini runs more than twice as fast as GPT-5 mini. Speed is not just comfort: in an agent loop, latency per call multiplies by the number of turns.

What the price gap buys back

OpenAI prices GPT-5.4 mini at $0.75 per million input tokens, $0.075 for cached input and $4.50 per million output, and OpenRouter lists the same rates at the time of writing. GPT-5.4 is $2.50, $0.25 and $15. Every rate is 3.3 times lower.

Put that against the benchmark gap. Losing roughly three points of SWE-Bench Pro to save 70% of the bill is a good trade when tasks are retryable and checked by tests, and a poor one when a single failure is expensive or invisible. The arithmetic that decides it is cost per completed task, not cost per token: a mini attempt costs about 30% of a GPT-5.4 attempt, so a mini-first setup that escalates failures to GPT-5.4 only loses money once mini fails on roughly 70% of tasks. In practice the extra latency and review effort of escalations bite well before that point. Model routing and fallbacks covers running mini first and escalating on failure, which usually captures most of the saving without the risk.

Mini as the subagent model

OpenAI positions mini as its strongest mini model for coding, computer use and subagents, and the API profile backs that up. In the Responses API it supports function calling, web search, file search, tool search, image generation, code interpreter, hosted shell, apply_patch, skills, computer use and MCP. It is also on Chat Completions and Batch.

That tool list is what makes it a good worker in an orchestrator-worker design. A larger model plans and reviews; several mini instances do the bounded pieces in parallel: reading a module and summarising it, writing tests for one file, checking a page renders, applying a patch and running the suite. Each piece is small enough that mini's few points of benchmark deficit rarely bite, and the parallelism makes its speed count twice. Subagents and delegation covers how to size those pieces.

400K context, 272K of input

The catalogue lists a 400,000-token context window, which is correct but easy to misread. OpenAI caps input at 272,000 tokens and reserves up to 128,000 for output. If you are moving work down from GPT-5.4, which has a 1.05M window, check that your largest prompts fit within 272K before you switch.

Knowledge cutoff is 31 August 2025, the same as the rest of the family. Reasoning effort runs from none, the default, through low, medium, high and xhigh. Benchmarks are usually reported at high settings, so set the level explicitly when you evaluate rather than comparing a default run with a vendor chart.

Mini, nano or full GPT-5.4

A decision rule that holds for most teams:

• Pick mini for subagents, routine coding, code review passes, computer-use steps and anything high-volume where a test or a check catches mistakes.
• Pick nano when the task is classification, extraction or ranking with a fixed output shape, and computer use is not involved. It is about a quarter of mini's input price, but its OSWorld-Verified score is 39.0% against mini's 72.1%. The GPT-5.4 nano guide covers where that gap matters.
• Pick full GPT-5.4 for planning, long-context work beyond 272K, ambiguous tasks where a failure is costly, and the reviewer seat in an orchestrator design. The GPT-5.4 guide covers its long-context pricing.

Snapshots and the batch endpoint

Pin the dated snapshot, gpt-5.4-mini-2026-03-17, in production so an alias update cannot change behaviour under you. OpenRouter also lists openai/gpt-5.4-mini:batch, the same model on an asynchronous endpoint at half price: $0.375 input and $2.25 output per million tokens at the time of writing. For overnight evaluation runs and bulk processing it makes an already cheap model cheaper still.

The practical takeaway: make mini the default worker, keep GPT-5.4 for the decisions, and measure cost per completed task for a week before you decide the split.