OpenAI launches GPT-5.4 for work and agents

OpenAI has launched GPT-5.4 in ChatGPT, the API and Codex, with a clear focus: AI should not just answer, but complete professional tasks from start to finish. The model combines reasoning, coding, tool use and work with documents, spreadsheets and presentations.

GPT-5.4 Pro is also arriving, aimed at especially complex tasks. In ChatGPT, the model appears as GPT-5.4 Thinking and can show an initial plan before it starts working, so you can adjust its direction without waiting for it to finish.

AI that works inside your applications

The main new feature is computer use. GPT-5.4 can interact with websites and software through screenshots, a keyboard and mouse, or libraries such as Playwright. In practice, an agent could navigate a portal, extract information, fill out forms and verify the result.

OpenAI says GPT-5.4 achieved a 75% success rate on OSWorld-Verified, a test of tasks in desktop environments, compared with 47.3% for GPT-5.2. That figure also exceeds the 72.4% recorded for people in that specific evaluation.

The model supports up to 1 million tokens of context, meaning it can keep much more information available during long tasks. This is useful for reviewing large volumes of documents, following a project that lasts several hours or coordinating many steps without losing track.

That does not mean every ChatGPT user will automatically get an autonomous agent capable of handling their entire computer. These features depend on how developers integrate them and on the safety confirmations they configure.

Improvements for professional work

GPT-5.4 is especially focused on tasks that typically consume hours of work:

• Creating and editing spreadsheets, presentations and documents.
• Analyzing contracts and preparing reports.
• Writing, testing and debugging code.
• Researching questions on the web that require consulting many sources.
• Coordinating multiple services through tools and connectors.

In an internal financial modeling test, OpenAI achieved an average score of 87.3%, compared with 68.4% for GPT-5.2. In another evaluation, participants preferred presentations generated by GPT-5.4 68% of the time over those made with the previous version.

In GDPval, a test covering knowledge work across 44 professions, GPT-5.4 matched or exceeded the performance of professionals in 83% of comparisons, compared with 70.9% for GPT-5.2. These results were published by OpenAI and come from controlled evaluations, so they do not guarantee the same performance in every real-world case.

Fewer errors and better tool use

OpenAI says GPT-5.4 is its most accurate model to date. In a set of queries whose errors had been flagged by users, its individual claims were 33% less likely to be false than those of GPT-5.2, while its complete answers were 18% less likely to contain at least one error.

The API also adds tool search. Instead of loading the descriptions of thousands of tools from the start, the model can search for the definition it needs just when it is about to use it. In a test with 36 MCP servers, this feature reduced total token use by 47% without changing accuracy.

For you, this could mean faster and cheaper agents when they work with many connected services. It also reduces the unnecessary information the model has to process in each request.

What changes in ChatGPT, Codex and the API

GPT-5.4 Thinking is available today to Plus, Team and Pro users, replacing GPT-5.2 Thinking as the main option. GPT-5.2 Thinking will remain in the legacy models section for three months, until June 5, 2026. Enterprise and Edu plans can enable early access from the administration settings.

GPT-5.4 Pro is reserved for Pro and Enterprise plans in ChatGPT, in addition to being available through the API. In the API, the models are identified as gpt-5.4 and gpt-5.4-pro.

The price of gpt-5.4 is $2.50 per million input tokens and $15 per million output tokens. It is more expensive than GPT-5.2, although OpenAI says it needs fewer tokens to solve many tasks. Codex's fast mode can increase generation speed by up to 1.5 times without changing the model or its capabilities.

The arrival of GPT-5.4 points to an important transition: the value of these models is no longer just in drafting an answer, but in planning, using software, checking results and completing long workflows. What remains to be seen is how much of that promise holds up outside OpenAI's tests and how much human oversight is used when these agents are deployed.