GPT-5.4 Mini and Nano: Smaller Models That Change How We Use AI

Mar 27, 2026 · #openai #gpt 5.4 #gpt 5.4 mini #gpt 5.4 nano #chatgpt #ai #artificial intelligence #ai models #subagent

Nearly as capable as the flagship, over twelve times cheaper, and designed to work as a team. OpenAI is changing what an AI model is supposed to be.

AI neural network visualization

On March 17, 2026, OpenAI released two new models: GPT-5.4 Mini and GPT-5.4 Nano. These are smaller, faster, and cheaper variants of the flagship GPT-5.4, which debuted earlier in March.

I covered this in AITU #01 — my weekly AI and tech series in Polish — but here I’m going deeper with full context and sources.

What These Models Are

GPT-5.4 Mini is a mid-size model designed for tasks where response speed directly shapes the user experience — coding assistants, multimodal tools, and systems that interpret what you see on screen.

GPT-5.4 Nano is the smallest and cheapest model in OpenAI’s lineup, built for high-volume simple tasks — sorting, extracting data from documents, classification, and lightweight coding support.

A simple way to think about it: imagine a company. There’s a director who makes strategic decisions. And there are employees who do the actual work — searching through documents, filling out forms, checking data. OpenAI is saying: do the same thing with AI. The big model thinks and plans. The smaller models execute.

Performance — Surprisingly Close to the Flagship

On SWE-Bench Pro — a test where the model receives real software engineering tasks from GitHub and has to solve them — Mini scored 54.4%. The flagship GPT-5.4 scored 57.7%. The gap is just 3.3 percentage points.

On OSWorld-Verified — a test measuring whether a model can “use a computer” (clicking, navigating interfaces, interpreting screenshots) — Mini scored 72.1% compared to the flagship’s 75%.

Nano is weaker — 39% on OSWorld-Verified — so it’s not suited for complex tasks. But on simpler work (sorting, classification, tool calling) it outperforms the older GPT-5 Mini.

Pricing

Mini: $0.75 per million input tokens, $4.50 per million output tokens.
Nano: $0.20 per million input tokens, $1.25 per million output tokens.

To put that in everyday terms — Nano can process roughly several hundred pages of text for twenty cents. The flagship GPT-5.4 costs $2.50 per million input tokens. Nano is over twelve times cheaper.

The Subagent Architecture — This Is the Interesting Part

Benchmarks and pricing are one thing. But the most important part of this release is how OpenAI talks about how to use these models.

The company explicitly positions Mini and Nano as subagents. In this approach, the big model (GPT-5.4) acts as a coordinator — it sets the goal, breaks it into subtasks, and delegates. The smaller models execute those subtasks in parallel — searching code, reviewing files, processing documents.

Within OpenAI’s Codex platform (their AI coding tool), Mini uses only 30% of the GPT-5.4 quota. That means a developer can handle simpler coding tasks at roughly one-third of the cost.

This is the same trend that was visible a week earlier at NVIDIA GTC — Jensen Huang talked about AI agents and systems where AI works as a team of collaborating models, not as one giant brain.

Where It’s Available

Mini works in the OpenAI API, in Codex, and in ChatGPT. Free and Go tier users can access it through the Thinking feature. It supports a 400K context window with text and image inputs.
Nano is available through the API only — it’s a tool for developers, not for the average user.

Why This Matters

GPT-5.4 Mini and Nano signal that the era of “one model that does everything” is ending. The future of AI is ecosystems — different models with different sizes, costs, and specializations, working together.

For everyday users, this means faster and cheaper AI services.
For developers, the ability to build systems where AI delegates tasks to other AI.
For businesses, real savings at scale.

And notably — Mini is on ChatGPT’s free plan. You can test it right now without paying for anything.