Reading · Benjamin Crozat's blog posts

What this GPT-5.4 nano guide covers

OpenAI introduced the GPT-5.4 family on March 5, 2026 in its GPT-5.4 launch post. The current GPT-5.4 nano model page places nano in that family, and the current pricing page puts it at the low end of that lineup.

That makes GPT-5.4 nano the model I would reach for when the job is small, repetitive, and high-volume: classification, extraction, ranking, and lightweight routing to sub-agents.

This guide uses the pinned snapshot gpt-5.4-nano-2026-03-17.

By the end, you will have:

a first successful GPT-5.4 nano Responses API call
a strict JSON routing workflow that can send work to the right sub-agent

If you want the fuller GPT-5.4 picture first, start with my GPT-5.4 API guide. If you are coming from the older family, compare this with GPT-5 nano.

Get your API key ready

You need an OpenAI account, a funded API project, and an API key from the API keys page. Keys are shown once, so save yours right away.

Then export it in your terminal.

macOS and Linux:

export OPENAI_API_KEY="sk-..."

Windows Command Prompt:

setx OPENAI_API_KEY "sk-..."

If you use setx, open a new terminal before testing the key.

Send your first GPT-5.4 nano request

GPT-5.4 nano uses the same Responses API shape as the rest of the GPT-5.4 family, so the first request stays simple:

curl -s https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.4-nano-2026-03-17",
    "input": [
      {
        "role": "user",
        "content": [
          { "type": "input_text", "text": "Say hello in one short sentence." }
        ]
      }
    ],
    "text": {
      "verbosity": "low"
    },
    "max_output_tokens": 80
  }'

If your key and billing are set up correctly, you should get a short greeting back.

Build something useful: route work to sub-agents

Nano is not where I would start for a long, nuanced response draft. It is where I would start for short, repeatable jobs that need a clean handoff.

Imagine your app receives this instruction:

Check the pricing copy, fix the headline if needed, and flag anything risky.

You want GPT-5.4 nano to:

classify the request
choose the right sub-agent
set a priority
return a short handoff note

That is a very good fit for nano because the output is narrow, structured, and easy to validate.

Return strict JSON with a schema

curl -s https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.4-nano-2026-03-17",
    "instructions": "You route small tasks to the right sub-agent for a SaaS app. Be cautious. Do not promise changes that were not requested. Keep handoff_note to 2 short sentences.",
    "input": [
      {
        "role": "user",
        "content": [
          {
            "type": "input_text",
            "text": "Check the pricing copy, fix the headline if needed, and flag anything risky."
          }
        ]
      }
    ],
    "text": {
      "verbosity": "low",
      "format": {
        "type": "json_schema",
        "name": "sub_agent_routing",
        "schema": {
          "type": "object",
          "properties": {
            "destination_agent": {
              "type": "string",
              "enum": ["copy_agent", "seo_agent", "legal_agent", "general_agent"]
            },
            "priority": {
              "type": "string",
              "enum": ["low", "medium", "high"]
            },
            "needs_human": {
              "type": "boolean"
            },
            "handoff_note": {
              "type": "string"
            }
          },
          "required": ["destination_agent", "priority", "needs_human", "handoff_note"],
          "additionalProperties": false
        },
        "strict": true
      }
    },
    "max_output_tokens": 200
  }'

That kind of output is easy to plug into a dispatcher:

destination_agent picks the next sub-agent
priority changes queue order
needs_human can stop unsafe automation
handoff_note gives the next model a clean starting point

If you want to move the same pattern into PHP afterward, my guide on using OpenAI’s API in PHP with openai-php/client picks up right there.

Why GPT-5.4 nano is different

GPT-5.4 nano is not just a smaller model. It is the one I would choose when cost and throughput matter more than broad reasoning.

OpenAI’s pricing page puts it at the low end of the GPT-5.4 family, which is why it makes sense for:

classification
data extraction
ranking
routing to sub-agents
preprocessing before a stronger model sees the hard cases

That also makes it a natural step up from the older GPT-5 nano guide when you want the newer GPT-5.4 family without jumping all the way to full GPT-5.4.

When GPT-5.4 nano is the right model

Pick GPT-5.4 nano when you care about:

very high request volume
low latency
low cost
strict structured output
lightweight routing before a bigger model or a human steps in

If the task starts needing broader judgment, GPT-5.4 is the more capable next stop. If the task stays simple but you want the older family, GPT-5 nano is still a useful comparison point.

Common mistakes with GPT-5.4 nano

1. Expecting it to behave like the flagship model

Nano is great at small, repeatable tasks. It is not where I would start for open-ended analysis.

2. Using long reply-drafting workflows by default

Nano works best when the output is short and predictable.

3. Forgetting to route hard cases onward

Nano is strongest when it filters, labels, or dispatches. It does not need to do everything itself.

If GPT-5.4 nano looks close to what you need, these are the next reads I would keep open:

GPT-5.4 nano API quick start with a real workflow