OpenAI released GPT-5.4 on March 5, 2026, describing it as its most capable model yet for professional work. The release adds native computer use, a 1 million token context window and, according to the company, a measurable drop in factual errors compared with its predecessor.
The model ships in three versions. The standard GPT-5.4 is intended for everyday use, GPT-5.4 Thinking allocates more compute to step-by-step reasoning on harder problems, and GPT-5.4 Pro is the top tier aimed at developers, researchers and professional users with the most demanding workloads. The tiering mirrors the structure OpenAI has used since GPT-5.
A 1 million token context
The headline specification is a context window of 1 million tokens, the largest OpenAI has offered. A context window is the amount of text a model can take into account at once; at 1 million tokens, the model can hold several long books or a large codebase in a single session. Google’s Gemini models have offered contexts of this size since 2024, so the change brings OpenAI to parity on that measure rather than ahead of it.
Large contexts matter for tasks such as reviewing lengthy contracts, analyzing full code repositories or working through long multi-step agent sessions without losing track of earlier instructions.
Native computer use
GPT-5.4 is the first OpenAI model with computer use built in rather than offered through a separate agent product. The model can view a screen, move a cursor, click and type to carry out tasks in ordinary software. OpenAI had previously delivered similar functionality through its Operator agent and later ChatGPT agent features; building the ability into the base model is meant to make it available to developers through the API.
Anthropic introduced a comparable computer use capability for Claude in late 2024, and the feature has since become a focus of competition among frontier labs as they push toward agents that complete tasks rather than just answer questions.
Accuracy claims
OpenAI says GPT-5.4 produces 18 percent fewer factual errors than GPT-5.2 on its internal evaluations. The figure comes from the company’s own testing and has not been independently replicated at the time of writing. OpenAI also published benchmark results on coding, reasoning and computer-use tasks, and said the model was tuned for finance and other enterprise work, releasing a ChatGPT for Excel add-in on the same day.
Competitive context
The launch comes as OpenAI’s annualized revenue has been reported to exceed $25 billion and as competition with Anthropic, Google and others intensifies. Anthropic released its own marketplace for Claude-based enterprise software the following day, and both companies are increasingly targeting the same corporate customers with agentic tools. GPT-5.4 is available in ChatGPT for paying subscribers and through OpenAI’s API, with the Pro tier limited to the company’s highest-priced plans.
