OpenAI brings GPT-5.4 mini to Free & Go users as nano goes API-only
On Wednesday, OpenAI expanded its GPT-5.4 lineup with the release of GPT-5.4 mini and GPT-5.4 nano. In ChatGPT, the broader family also includes GPT-5.4 Thinking and GPT-5.4 Pro, while Instant is still the fast mode associated with GPT-5.3 Instant.
While Mini is the smaller model, it’s meant to do the heavier lifting without pushing the cost too high. OpenAI says it beats GPT-5 mini across coding, reasoning, and tool use, while running more than twice as fast. It comes with a 400,000-token context window, costs $0.75 per 1 million input tokens and $4.50 per 1 million output tokens, and is available in the API, Codex, and ChatGPT.
And Nano gets the same 400,000-token context window, but OpenAI is selling it on price more than muscle. It is API-only, starts at $0.20 per 1 million input tokens and $1.25 per 1 million output tokens, and is aimed at simpler repeated work.
Benchmarks for Comparing GPT-5 mini, GPT-5.4 nano, and GPT-5.4 mini
The quickest way to see what changed is to look at how the new pair stacks up against GPT-5 mini on coding and tool-heavy work.
| Benchmark | GPT-5 mini | GPT-5.4 nano | GPT-5.4 mini |
| SWE-Bench Pro | 45.7% | 52.4% | 54.4% |
| Terminal-Bench 2.0 | 38.2% | 46.3% | 60.0% |
| MCP Atlas | 47.6% | 56.1% | 57.7% |
| Toolathlon | 26.9% | 35.5% | 42.9% |
| τ2-bench telecom | 74.1% | 92.5% | 93.4% |
What the scores suggest
The table makes the hierarchy easier to see. Nano moves up meaningfully from GPT-5 mini on coding fixes and tool use while keeping the low-price pitch. Mini then pulls further ahead, especially once the work starts to look more like real developer or agent workflow. The Terminal-Bench jump suggests mini is better at the messy part of software work, where the model has to deal with commands, state, and the shape of an actual task, not just produce a neat answer.
That is also why mini matters more and OpenAI is pitching mini as the lower-cost model that can still do serious work. The company says mini is built for coding, computer use, and in Codex it can take narrower subtasks while bigger models handle planning and final judgment.
Then why Nano
Nano may be the quieter part of the launch, but it is also the more revealing one. OpenAI is giving nano a narrower lane and pricing it so that lane becomes attractive. Classification, data extraction, ranking, subagents, repeated utility work — that is not glamorous, but it is the kind of work that shows up everywhere once teams start automating anything at scale.
Also read: Google, Amazon, Meta, Microsoft, and OpenAI are among 11 companies that signed a new anti-scam accord aimed at tackling cross-platform fraud.
Benchmarking without the hype
Benchmarks can show where a model improved, but they can only take you so far. They do not tell you how that model will hold up once it is dropped into the middle of actual work, where tasks are messier, tools break, and instructions are rarely clean. What OpenAI is showing here is a clearer split at the cheaper end of its lineup: Nano is built for repeated utility work. Mini is the one meant to handle tougher tasks without pushing the cost too high.
The takeaway
What changed with GPT-5.4 mini and nano is more than that the smaller models got better. OpenAI is making the lower end of its lineup less blurry. Mini is the cheaper model it wants closer to codebases, tools, and screens. Nano is the cheaper model to run repeated utility work at scale. That is the point of the release.
Not one more small model, but two smaller models with different jobs from the start.
Y. Anush Reddy is a contributor to this blog.
