OpenAI launches GPT-5.4 mini and nano: faster, but more expensive

What is new

On March 17, 2026, OpenAI expanded its GPT-5.4 family with two smaller models: GPT-5.4 mini and GPT-5.4 nano. They are built for tasks where response time matters, such as coding help that needs to feel fast, small helper programs that handle subtasks, and systems that read screenshots.

Mini can be used in OpenAI's developer tool (API), in the coding assistant Codex and in ChatGPT. Nano is the smallest version and is only available via the API, meaning for developers and businesses, not as a button in ChatGPT. Both models can remember up to 400,000 tokens (pieces of text) at a time, equivalent to a very long book.

What is clever

GPT-5.4 mini runs more than twice as fast as its predecessor GPT-5 mini and comes close to the large GPT-5.4 in several tests. On SWE-Bench Pro, which measures the ability to fix real programming bugs, mini scores 54.4 percent versus 57.7 for the large model. On OSWorld-Verified, where the AI has to control a computer, the figure is 72.1 versus 75.0 percent. On Terminal-Bench 2.0, the gap is bigger: 60.0 versus 75.1 percent.

Nano is for light, repetitive tasks such as sorting, extracting data from text, ranking results and simple coding help. According to The Decoder, it scores 52.4 percent on SWE-Bench Pro, but that figure should be taken with caution, as it is close to mini's.

Cheaper or better?

Better, yes. Cheaper, no, not compared with the old models. Developer prices per million tokens are:

- GPT-5.4 mini: $0.75 for input and $4.50 for output - GPT-5.4 nano: $0.20 for input and $1.25 for output

That is about 3 times more expensive on input and 2.25 times more expensive on output than GPT-5 mini. For nano, it is 4 times more expensive on input and about 3 times more expensive on output than GPT-5 nano. But they are still much cheaper than the large models, and you get more for your money than before. I have not been able to confirm comparisons with competitors' small models from Google and Anthropic in the sources.

For ordinary users, there is no extra bill, since mini is included in ChatGPT. Exactly what it costs per month or which subscriptions get access is not stated in the sources I have been able to read. Nor has anything been said about special terms for Denmark or the EU.

What it is good at

Mini suits programming where you want fast answers without paying for the large model, agents that control a computer, and apps that need to understand images in real time. It is also an obvious helper that takes the small subtasks while a larger model plans.

Nano is for companies with large volumes of simple tasks, such as sorting inquiries, extracting information from documents or ranking search results, where low price and speed count more than depth.