GPT-5.4

// Overview

GPT-5.4 is the flagship large language model OpenAI released on March 5, 2026, in the GPT (Flagship / Thinking) line. It launched first as GPT-5.4 Thinking and GPT-5.4 Pro — neither available to free-tier users — and was followed on March 17, 2026 by the smaller GPT-5.4 mini (free tier) and GPT-5.4 nano (API only).

OpenAI positions GPT-5.4 as its most capable and efficient frontier model for professional work. The headline feature is built-in computer use: the model can operate desktop environments natively rather than relying on a bolted-on agent layer. It also ships a 'tool search' mechanism that OpenAI says cuts token usage substantially in tool-heavy workflows without hurting accuracy, and OpenAI reports the model solves comparable problems using fewer tokens than GPT-5.2.

On reliability, OpenAI reports GPT-5.4 is 33% less likely than GPT-5.2 to make errors in individual factual claims and 18% less likely to produce a response containing any error. The API exposes a context window of up to 1 million tokens — the largest OpenAI has offered — with standard pricing of $2.50 per million input tokens and $15.00 per million output tokens.

// Benchmarks

1. OSWorld-Verified (computer use)75%
2. GDPval (knowledge work)83%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Standard tier (GPT-5.4). Requests above ~272K context move to long-context rates. Smaller variants are cheaper: GPT-5.4 mini ($0.75/$4.50) and GPT-5.4 nano ($0.20/$1.25).

Pricing source ↗

// Strengths

• Native computer use for operating desktop apps and browsers without a separate agent harness
• 1M-token context window — the largest OpenAI offers — for large codebases and document sets
• Improved token efficiency, including a tool-search mechanism that reduces cost in tool-heavy workflows
• Lower factual error rate than GPT-5.2 (33% fewer claim-level errors per OpenAI)
• Strong professional/knowledge-work performance, scoring 83% on OpenAI's GDPval evaluation

// Best for

• Agentic coding and software engineering over large repositories
• Computer-use automation: operating desktop and web applications end to end
• Deep research and long-document analysis using the 1M-token context
• Tool-heavy and multi-step professional workflows where token efficiency matters
• High-stakes knowledge work that benefits from a lower factual error rate

// How to access

// GPT (Flagship / Thinking) — every version

The full lineage of the GPT (Flagship / Thinking) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.