OpenAI GPT-5.4 Documentation

GPT-5.4 is OpenAI's flagship frontier model that unifies advanced reasoning, coding, and autonomous workflows into a single system. Absorbing the capabilities of specialist engineering branches directly into its mainline architecture, it acts as a highly integrated, general-purpose engine optimized for complex document synthesis, multi-file code generation, and multi-step agentic orchestrations.

Key Capabilities

Absorbed Engineering & Codex Mastery: Natively incorporates top-tier programming performance from separate developer variants, allowing developers to execute extensive code maintenance, refactoring, and test-suite generation out of the box.
Native Computer Use (CUA): OpenAI's first mainline model with native, state-of-the-art computer-use capabilities. It scores $75\%$ on the OSWorld-Verified benchmark, allowing autonomous sub-agents to interface cleanly with desktop environments and cross-application workflows.
Million-Token API Window: Supports an expansive 1,050,000 token input context window via the API, letting developers pass complete medium-sized repositories or dozens of long telemetry streams concurrently.
128,000 Maximum Output Target: Features an ultra-wide output ceiling (128K tokens) that completely eliminates text truncation issues during continuous large-scale generation and script building.
Upfront Planning Visibility: Integrates structured reasoning traces that output initial step-by-step action plans, enabling developers to monitor or intercept long-horizon agent execution paths before completion.

Request Parameters

To interact with this flagship model via the OpenAI Responses API or a compatible network gateway:

Parameter	Type	Required	Description
`model`	`string`	Yes	Use `"gpt-5.4"`.
`messages`	`array`	Yes	Array of standard conversation objects supporting text and multi-modal visual payloads.
`reasoning_effort`	`string`	No	Directs thinking token allocation boundaries. Supports `"low"`, `"medium"`, or `"high"`.
`max_completion_tokens`	`integer`	No	Controls output limits. Supports a maximum generation footprint of up to 128,000 tokens.

Reasoning Mode Configuration

gpt-5.4 supports internal reasoning via the reasoning_effort parameter. This controls how much compute the model dedicates to its hidden chain-of-thought before producing the final response.

Effort Levels

Value	Description
`"low"`	Minimal reasoning; fastest and cheapest
`"medium"`	Balanced reasoning depth and speed
`"high"`	Maximum reasoning; best for hard multi-step problems

Request Example

json Copy

{
  "model": "gpt-5.4",
  "reasoning_effort": "high",
  "messages": [
    {
      "role": "user",
      "content": "Your question here"
    }
  ]
}

Note: reasoning_effort is a top-level parameter in the request body. Not all GPT models support it — it applies to reasoning-capable variants only.

gpt-5.4

Input