# What are GPT-5.4 mini and GPT-5.4 nano, and why do small models matter?

Date: 2026-03-18 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-are-gpt-5-4-mini-and-nano/

**Answer:** GPT-5.4 mini and GPT-5.4 nano are smaller, faster and cheaper versions of GPT-5.4 that OpenAI released on March 17, 2026; mini comes close to the full model on several coding and computer-use tests, and free ChatGPT users can reach it through the Thinking option. Nano is the smallest and cheapest, offered only to developers through the API for simple, high-volume jobs such as sorting and extracting data.

Yesterday, Tuesday, March 17, 2026, OpenAI added two small members to the GPT-5.4 family it introduced earlier this month. **GPT-5.4 mini** and **GPT-5.4 nano** bring much of the larger model's skill to versions that answer faster and cost far less to run. OpenAI calls them its most capable small models so far.

The news that matters to most people: mini is now within reach of ChatGPT's Free and Go users. GPT-5.4 itself was aimed at paying professionals, as Engadget points out; mini lets everyone try a slice of it.

## What OpenAI released

A **small model** is usually built with fewer internal settings, or parameters, than a flagship model. That makes it quicker and cheaper, usually at some cost in accuracy on the hardest problems. OpenAI published these scores, with each model at its highest reasoning setting:

| Test (what it measures) | GPT-5 mini (older) | GPT-5.4 nano | GPT-5.4 mini | GPT-5.4 |
|---|---|---|---|---|
| SWE-Bench Pro (fixing real code) | 45.7% | 52.4% | 54.4% | 57.7% |
| OSWorld-Verified (using a computer) | 42.0% | 39.0% | 72.1% | 75.0% |
| GPQA Diamond (expert science questions) | 81.6% | 82.8% | 88.0% | 93.0% |

Mini lands close to the full model on coding and computer use, and OpenAI says it runs more than twice as fast as GPT-5 mini, the small model it replaces. Nano trails on computer use but holds up well on coding and science questions for its size.

Where you can use them:

| | GPT-5.4 mini | GPT-5.4 nano |
|---|---|---|
| ChatGPT | Free and Go: choose "Thinking" in the + menu. Paid plans: used as backup when you hit your GPT-5.4 Thinking limit | Not available |
| Codex | App, command line, editor extension and web; uses 30% of GPT-5.4's quota | — |
| API price per million tokens (input / output) | $0.75 / $4.50 | $0.20 / $1.25 |
| API context window | 400,000 tokens | — |

A **token** is a small piece of text, often part of a word; the **context window** is how much text a model can consider at once.

## How small models fit into bigger systems

The most interesting idea in the launch is not a score but a way of working: **subagents**. An agent is an AI that carries out a multi-step task with tools. A subagent is a helper that a main agent hands a narrow piece of the job to.

OpenAI describes this setup in Codex, its coding agent. The large GPT-5.4 plans the work and makes final decisions, while several GPT-5.4 mini helpers run side by side: one searches the codebase, another reads a long file, a third goes through documentation. Think of a head chef who designs the menu and tastes every plate, while line cooks chop and sauté in parallel. The chef's time is expensive; the cooks keep the kitchen fast.

Nano suits even simpler, repetitive work. OpenAI recommends it for **classification** (putting items into categories), **data extraction** (pulling names, dates or amounts out of text), **ranking** results, and light coding helpers.

Mini can also read screenshots of busy interfaces quickly, which helps agents that operate a computer by looking at the screen.

## A worked example: sorting customer emails with nano

Imagine a small online shop that receives 2,000 emails a week. A developer could send each one to GPT-5.4 nano with a short, strict instruction:

```prompt
Classify this customer email into exactly one category:
ORDER_STATUS, RETURN, PRODUCT_QUESTION, COMPLAINT, OTHER.
Then extract the order number if there is one.
Reply only as: category | order number (or "none").
```

1. Nano labels each email in a fraction of a second, at low cost per message.
2. Emails marked COMPLAINT go to a person right away.
3. Only the tricky cases are passed to a larger model to draft a careful reply.

The cheap model does the high-volume sorting; the expensive one is saved for work that needs judgment.

> **Try it:** If you use ChatGPT for free, open the + menu, choose Thinking, and give GPT-5.4 mini a small multi-step task, such as turning a messy shopping list into a table grouped by store aisle. Note how quickly it answers.

## Why it matters

Most real AI use is not one brilliant answer but thousands of routine ones: tagging, summarizing, checking, searching. For those, speed and price decide whether a product is practical. Mini and nano make that routine work cheaper while staying close enough to the flagship that developers can mix sizes freely.

For everyday users, the change is about access. A model that approaches GPT-5.4 on several tests is now in the free tier, through the Thinking option. That narrows the gap between what paying and non-paying users can try.

## When should you choose a small model like GPT-5.4 mini or nano?

Choose **mini** when you want quick, capable help and do not need the very best answer: coding edits, short research, reading screenshots, or as the helper models inside an agent. In ChatGPT's free tier, it is the Thinking option to reach for when a question needs a few steps of reasoning.

Choose **nano** when you are building something that runs the same simple task many times: sorting messages, pulling fields from forms, ranking search results. Write tight instructions with a fixed output format, as in the example above, and test it on a sample of real data before trusting it.

Reach for the full **GPT-5.4** when the task is long, ambiguous or high-stakes, or when mistakes are costly. A practical rule is to start small, measure the errors, and move up a size only where the results show you need to.

## Sources

- Introducing GPT-5.4 mini and nano: https://openai.com/index/introducing-gpt-5-4-mini-and-nano/
- GPT-5.4 mini brings some of the smarts of OpenAI's latest model to ChatGPT Free and Go users: https://www.engadget.com/ai/gpt-54-mini-brings-some-of-the-smarts-of-openais-latest-model-to-chatgpt-free-and-go-users-170000585.html