What is GPT 5.4 Nano?
GPT 5.4 Nano is OpenAI's March 17, 2026 small model, the cheapest GPT-5.4-class model in the family. OpenAI built it for simple high-volume work: classification, data extraction, ranking, and coding sub-agents that carry the supporting tasks around a larger model. It runs on Unifically as openai/gpt-5.4-nano, accepts text and image input, and outputs text. The 400k-token context window and 128k-token max output match the bigger GPT-5.4 models, so a cheap call can still read a lot of input. Nano is API-only; there is no ChatGPT version.
Key features of GPT 5.4 Nano
400k token context window
Long document sets, big extraction batches, or a large codebase slice fit in one request, with a callable max output of 128k tokens.
Benchmark gains over the GPT-5 small models
At xhigh effort it scores 52.4% on SWE-Bench Pro, 46.3% on Terminal-Bench 2.0, and 82.8% on GPQA Diamond, ahead of GPT-5 mini on all three.
Best accuracy per dollar in the family
On SWE-Bench Pro cost curves, Nano returns more accuracy per dollar than GPT 5.4 Mini or the full GPT-5.4 at the same spend.
Five reasoning effort settings
Effort runs from none (the default) through low, medium, and high to xhigh. Simple lookups stay fast; hard tasks get more thinking time.
Vision, tools, and structured outputs
Image input covers scanned documents, screenshots, and charts. Function calling, structured outputs, and streaming are all supported, so it slots into agent stacks and returns clean JSON.
Best for
Classification
Ticket triage, moderation queues, and intent labeling at high volume.
Data extraction
Pulling fields from invoices, forms, and scanned pages with image input.
Ranking and routing
Search reranking and model routing where per-call cost decides the design.
Coding sub-agents
File summaries, small patches, and checks inside a larger coding agent.
High-volume pipelines
Batch jobs that run millions of calls and need predictable spend.
Low-latency features
Autocomplete, quick replies, and interactive tools at effort none or low.
Use cases
Build an extraction pipeline that reads scanned invoices and returns structured JSON through function calling. Use it as the router in front of larger models: Nano classifies each request, answers the easy ones, and passes the hard ones up to GPT 5.4. It fits coding agents as a sub-agent that summarizes files, drafts small patches, and runs checks while a bigger model plans. For search products, it makes a cheap reranker that scores candidates against a query. And with 400k tokens of context, one call can label or summarize a document set that would need chunking elsewhere.
Limitations
Computer use is the clear weak spot. GPT 5.4 Nano scores 39.0% on OSWorld-Verified, below even GPT-5 mini at 42.0%. Hand browser and desktop agents to GPT 5.4 Mini at 72.1% or the full GPT-5.4.
Top scores take time. The headline numbers come from xhigh effort, and on SWE-Bench Pro latency curves Nano runs longer than GPT 5.4 Mini to reach similar accuracy. xhigh runs are also verbose, so output-token spend climbs on hard tasks.
Input is text and images only, with no audio or video, and fine-tuning is not supported.
The knowledge cutoff is August 31, 2025. Anything after that date needs retrieval or web search in your stack.
GPT 5.4 Nano vs GPT-5 mini
GPT 5.4 Nano beats GPT-5 mini, one size class above the old GPT-5 nano, on almost every published benchmark: 52.4% vs 45.7% on SWE-Bench Pro, 46.3% vs 38.2% on Terminal-Bench 2.0, 35.5% vs 26.9% on Toolathlon, and 82.8% vs 81.6% on GPQA Diamond. The one loss is OSWorld-Verified at 39.0% vs 42.0%. In short: the new nano replaces the old mini for text, reasoning, and tool work, but not for computer use.
When to use GPT 5.4 Nano
Use GPT 5.4 Nano when call volume and cost drive the design: classification, extraction, ranking, routing, and sub-agent work inside larger agent stacks. Start at effort none for speed and step up to xhigh only when a task earns it. Move to GPT 5.4 Mini for computer use or when accuracy on hard coding and terminal tasks matters more than price, and to the full GPT-5.4 when one model has to do it all.
API examples
Call GPT 5.4 Nano from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model on the platform. Full parameter docs live at docs.unifically.com/models/llm/openai/gpt-5.4-nano.
curl -X POST https://api.unifically.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "openai/gpt-5.4-nano",
"messages": [
{ "role": "user", "content": "Classify this support ticket as billing, bug, or feature request: I was charged twice this month." }
]
}'
The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.
FAQs
People also ask
openai/gpt-5.4-nano, called through the OpenAI-compatible POST /v1/chat/completions endpoint with your Unifically API key.
400k tokens, with a callable max output of 128k tokens. That fits long document sets or a large codebase slice in one request.
Text and image input with text output. Scanned pages, screenshots, and charts work as input. There is no audio or video support.
For sub-agent work, yes. It scores 52.4% on SWE-Bench Pro and 46.3% on Terminal-Bench 2.0 at xhigh effort. Use it for supporting tasks like file summaries and small patches, and keep a larger model on the plan.
Yes. Reasoning effort runs from none (the default) through low, medium, and high to xhigh, so you pick the speed-accuracy trade per request.
Not well. It scores 39.0% on OSWorld-Verified, below GPT-5 mini's 42.0%. Route browser and desktop tasks to GPT 5.4 Mini or the full GPT-5.4.
No. Fine-tuning is not supported. Prompting, function calling, and structured outputs cover most tuning needs in practice.
August 31, 2025. Add retrieval or web search to your stack for anything newer.