GPT nano is OpenAI’s smallest and cheapest model tier, built for classification, data extraction, ranking, and sub-agents rather than conversation. The current one carrying the name is GPT-5.4 nano, released 17 March 2026.
Two things about it get almost no coverage. Its input pricing is exactly four times what GPT-5 nano charged. And the context window on this tier has gone down, not up, losing 60 percent of its capacity since the 4.1 generation.
The Whole Nano Line
| Model | Released | Context | Input / Output per million tokens |
|---|---|---|---|
| GPT-4.1 nano | 14 April 2025 | 1,000,000 | Superseded |
| GPT-5 nano | 7 August 2025 | 400,000 | $0.05 / $0.40 |
| GPT-5.4 nano | 17 March 2026 | Not published in the same form | $0.20 / $1.25 |
OpenAI’s own documentation for GPT-4.1 nano now carries a note recommending GPT-5 nano instead for more complex work, which is a polite way of retiring it.
The Cheap Tier Stopped Being Cheap
Run the numbers and the jump is stark. Input went from $0.05 to $0.20 per million tokens, exactly four times. Output went from $0.40 to $1.25, a little over three times.
OpenAI released GPT-5.4 mini and nano together, and both landed at roughly four times the price of their GPT-5 equivalents. For most people that is irrelevant, because nano-tier spending is measured in cents. For anyone running classification over millions of records, it quadrupled a line item overnight.
The models are better, and OpenAI describes 5.4 nano as a significant upgrade over its predecessor. But the framing matters. This is not the budget tier getting cheaper as models improve, which is the pattern everyone expects. It is the budget tier moving upmarket while a new floor gets defined somewhere else.
Worth flagging if you have regional requirements: data residency endpoints carry a 10 percent uplift on GPT-5.4 nano. Small, but it compounds at the volumes this model is designed for.
Context Went Backwards
GPT-4.1 nano shipped with a one million token context window. It was the headline feature, and OpenAI made a point of it at launch alongside scores of 80.1 percent on MMLU and 50.3 percent on GPQA.
GPT-5 nano dropped that to 400,000.
Nobody announced this as a reduction because it was framed as a new model rather than a downgrade, but if you built something on 4.1 nano that relied on stuffing very long documents into a cheap model, the successor cannot do it. That is the sort of regression that surfaces in production rather than in a launch post.
What GPT-5.4 Nano Is Actually For
OpenAI is specific, and the list is narrower than people assume: classification, data extraction, ranking, and coding subagents handling simpler supporting tasks.
Notice what is absent. No general chat, no complex reasoning, no user-facing assistant work. Nano is infrastructure. It is the model you call ten thousand times an hour to decide whether an email is a complaint, not the one you put in front of a customer.
One genuinely new capability is worth knowing about. GPT-5.4 nano supports reasoning effort settings running from none, which is the default, through low, medium, high and xhigh. A budget model with a dial for how hard it thinks is unusual, and it means the same model can serve a cheap high-volume path and a slower, more careful one without switching models.
That is the same idea behind the efficiency work we covered in SWE-1.6, where the gains came from making a model think less wastefully rather than making it bigger.
The Nano Name May Be On Its Way Out
Here is the part that will date every article written about this before now.
OpenAI’s current generation is the GPT-5.6 family, and it does not use size names. The three variants are Sol, Terra and Luna, with Luna occupying the cost-efficient slot that nano and mini used to share. Above them sits GPT-6 Astra.
No nano variant has been announced under the 5.6 naming. So GPT-5.4 nano remains the most recent model carrying the name, and the tier it represented now goes by a codename that tells you nothing about its size or price.
Whether that is a permanent shift or a one-generation experiment is not something anyone outside OpenAI can say. But if you are searching for GPT nano and finding nothing newer than March, that is why.
Where You Can Actually Use It
Availability is documented inconsistently and worth checking before you plan around it.
OpenAI’s own launch announcement states that GPT-5.4 mini and nano are available in the API, Codex, and ChatGPT. Other documentation describes nano as API-only, with mini being the variant that reached free-tier ChatGPT users.
The likeliest reading is that the announcement covers both models loosely and nano lives in the API and Codex rather than the chat interface. Confirm against the model picker in whichever product you are using rather than trusting either summary.
Third-party gateways also serve it, which introduces its own variables around pricing, version pinning and data handling. Our guide to AI model aggregators and gateways covers what changes when you route through an intermediary instead of calling OpenAI directly.
Which One Should You Use
Pick GPT-5.4 nano if you need the current version of this tier and the four-times price change is still trivial at your volume. For most developers it is, and the capability gain is real.
Stay on GPT-5 nano if your workload is genuinely simple and you are running it at scale. At $0.05 per million input tokens it remains dramatically cheaper, and classification does not need frontier reasoning. Test both against your actual data before assuming newer is better here.
Look at the 5.6 family instead if you want a model with support going forward rather than one from a generation OpenAI has already moved past. Luna is where the cost-efficient work now sits.
Do not use nano at all if your task involves multi-step reasoning, long documents, or anything a user will read directly. That is what the larger tiers exist for, and our roundup of the best AI for coding covers the models built for work nano will fail at.
Frequently Asked Questions
What is the latest GPT nano model?
GPT-5.4 nano, released 17 March 2026. The GPT-5.6 generation that followed uses codenames rather than size labels, and no nano variant has been announced under it.
How much does GPT-5.4 nano cost?
$0.20 per million input tokens and $1.25 per million output tokens, with a 10 percent uplift on data residency endpoints.
Is nano available in ChatGPT?
Documentation is inconsistent. OpenAI’s launch post lists API, Codex and ChatGPT for both mini and nano, while other sources describe nano as API-only. Check the model list in your own account.
The Practical Call
Nano is a components tier, not a product you sit and talk to. If you are building something that makes thousands of small judgements, this is the right shelf to shop from, and the reasoning effort dial makes GPT-5.4 nano more flexible than anything that came before it on this tier.
Just do the arithmetic before you migrate. A four-times input price increase is invisible at a hundred requests a day and painful at ten million a month, and the older model is still there. If you are new to how these tiers relate to each other, our explainer on what GPT actually stands for is the place to start.
Published: September 10, 2026
