GPT-5.4 API quick start with a real workflow

What this GPT-5.4 guide covers

OpenAI released GPT-5.4 on March 5, 2026 in its GPT-5.4 announcement. The current GPT-5.4 model page positions it as the frontier model for complex professional work, and the broader models guide recommends it as the default place to start if you are not sure which GPT-5 model to pick.

This guide is about the full GPT-5.4 model, not GPT-5.4 Pro. Every main example below uses the pinned snapshot gpt-5.4-2026-03-05.

By the end, you will have:

• a first successful GPT-5.4 Responses API call
• a structured support-triage workflow
• a working example of GPT-5.4’s phase-aware assistant messages

Get your API key ready

You need an OpenAI account, a funded API project, and an API key from the API keys page.

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 request

GPT-5.4 keeps the same clean baseline request shape as GPT-5.2, so the first request is easy:

curl -s https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.4-2026-03-05",
    "input": [
      {
        "role": "user",
        "content": [
          { "type": "input_text", "text": "Say hello in one short sentence." }
        ]
      }
    ],
    "reasoning": {
      "effort": "none"
    },
    "text": {
      "verbosity": "low"
    },
    "max_output_tokens": 80
  }'

That exact request completed successfully for me and returned Hello!.

Build something useful: support triage

The first real workflow can stay simple.

Incoming message:

Hi, I was billed twice for my Pro plan today. Please refund the extra charge.

Goal:

1. classify the issue
2. set a priority
3. decide whether a human should step in
4. draft a safe reply

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-5.4-2026-03-05",
    "instructions": "You triage support messages for a SaaS app. Be cautious. Do not promise actions the billing team has not confirmed. Keep reply_draft to 2 short sentences.",
    "input": [
      {
        "role": "user",
        "content": [
          {
            "type": "input_text",
            "text": "Hi, I was billed twice for my Pro plan today. Please refund the extra charge."
          }
        ]
      }
    ],
    "reasoning": {
      "effort": "none"
    },
    "text": {
      "verbosity": "low",
      "format": {
        "type": "json_schema",
        "name": "support_triage",
        "schema": {
          "type": "object",
          "properties": {
            "category": {
              "type": "string",
              "enum": ["billing", "bug", "account", "feature_request", "other"]
            },
            "priority": {
              "type": "string",
              "enum": ["low", "medium", "high"]
            },
            "needs_human": {
              "type": "boolean"
            },
            "reply_draft": {
              "type": "string"
            }
          },
          "required": ["category", "priority", "needs_human", "reply_draft"],
          "additionalProperties": false
        },
        "strict": true
      }
    },
    "max_output_tokens": 220
  }'

That exact request completed successfully for me and returned JSON in this shape:

{
  "category": "billing",
  "priority": "high",
  "needs_human": true,
  "reply_draft": "Sorry about the duplicate charge. I'm escalating this to our billing team to review the charges and contact you about any applicable refund."
}

That is short, predictable, and easy to plug into real app logic.

GPT-5.4’s new practical parameter: phase

The most important GPT-5.4-specific addition for multi-turn agent work is the phase field on assistant messages.

OpenAI’s latest model guide uses phase to distinguish assistant commentary from the final answer, and GPT-5.4 also returned phase: "final_answer" in my live tests by default.

Here is a working example that sends prior assistant messages with phases, then asks GPT-5.4 to summarize what the assistant already said:

curl -s https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.4-2026-03-05",
    "input": [
      {
        "role": "assistant",
        "phase": "commentary",
        "content": [
          { "type": "output_text", "text": "I am checking the policy first." }
        ]
      },
      {
        "role": "assistant",
        "phase": "final_answer",
        "content": [
          { "type": "output_text", "text": "The policy allows a replacement within 30 days." }
        ]
      },
      {
        "role": "user",
        "content": [
          { "type": "input_text", "text": "Summarize the assistant's prior answer in 5 words." }
        ]
      }
    ],
    "reasoning": {
      "effort": "none"
    },
    "text": {
      "verbosity": "low"
    },
    "max_output_tokens": 40
  }'

That exact request worked for me and returned:

Replacement allowed within 30 days.

Why this matters:

• commentary is useful for progress-style assistant updates
• final_answer marks what the user should treat as the real answer
• multi-turn tool-heavy apps become easier to structure cleanly

This is one of the clearest ways GPT-5.4 feels more agent-ready than the earlier GPT-5 releases.

GPT-5.4 is also the long-context jump

The current model page lists a 1,050,000-token context window and 128,000 max output tokens. It also notes that prompts above 272K input tokens are priced differently: 2x input and 1.5x output for the whole session.

So yes, GPT-5.4 is the best GPT-5-family default right now, but it is also the first one where long-context cost planning becomes something you need to think about early.

Common mistakes with GPT-5.4

1. Treating it like GPT-5.2 with only a better benchmark score

GPT-5.4 also changes how tool-heavy and multi-turn interactions are modeled. The phase field is part of that story.

2. Ignoring the long-context pricing threshold

If you regularly go over 272K input tokens, the pricing model changes for the whole session.

3. Using a moving alias when you want stable tests

For tutorials, evals, and regression checks, prefer gpt-5.4-2026-03-05.

When full GPT-5.4 is worth using

Use GPT-5.4 when you want the newest frontier default for coding, professional work, long context, and agentic workflows.

The current model page lists:

• 1,050,000 context window
• 128,000 max output tokens
• $2.50 input and $15 output per 1M tokens
• reasoning.effort support for none, low, medium, high, and xhigh
• support for tools like web search, file search, code interpreter, hosted shell, apply patch, skills, computer use, MCP, and tool search

That is a powerful package, but it is not the cheapest one. If the task is simpler, GPT-5.1 or GPT-5.2 may be the more practical fit.

If you want the next useful comparisons from here, these are the ones I would keep open:

• Compare GPT-5.4 with the previous professional-work frontier model, GPT-5.2
• See the original GPT-5.0 API guide to understand where these controls started
• Move this same structured-output pattern into PHP
• Get a simpler mental model for how GPT-style systems behave before you build a bigger agent