OpenAI Launches GPT-5.4 — A Million-Token Context Window and Native Computer Use Arrive Together
GPT-5.4 ships with a 1M-token API context window, built-in computer-use capabilities, and a new Thinking mode that scores at expert level on economic benchmarks.
The Biggest Context Window in the Business
OpenAI officially launched GPT-5.4 on March 5, and it represents the most significant capability jump since GPT-4. The headline number is a one-million-token context window available through the API — roughly 50 to 100 times larger than what most developers have been working with. That means entire codebases, multi-hundred-page legal contracts, or months of conversation history can fit into a single prompt without chunking or summarization workarounds.
The standard context window ships at 272K tokens, with requests beyond that threshold counting at double the normal usage rate. But for enterprise workflows that involve massive document analysis, long-horizon agent planning, or multi-file code understanding, the million-token option changes what's architecturally possible. Developers no longer need to build elaborate retrieval pipelines just to give their models enough context to be useful.
Computer Use Goes Mainstream
Perhaps more consequential than the context expansion is GPT-5.4's native computer-use capability. This is OpenAI's first general-purpose model with built-in ability to operate desktop applications, navigate interfaces, click buttons, fill forms, and execute multi-step workflows across software — all without requiring external tool-use frameworks or custom integrations.
In the Codex environment and through the API, agents powered by GPT-5.4 can now carry out complex workflows that span multiple applications. The practical implications are enormous: automated QA testing, data entry across legacy systems, report generation that pulls from multiple dashboards — tasks that previously required either human operators or brittle RPA scripts.
Thinking Mode Hits Expert Level
GPT-5.4 ships in three variants: the standard model, GPT-5.4 Pro for high-performance workloads, and GPT-5.4 Thinking — a reasoning-optimized version that scored 83.0% on the GDPVal benchmark, placing it at or above the level of human experts on economically valuable tasks. The Thinking variant also incorporates the frontier coding capabilities from GPT-5.3-codex, making it particularly strong for software engineering workflows.
A new tool search feature rounds out the release. Instead of receiving the full definition of every available tool in the system prompt, the model gets a lightweight list and can look up detailed tool definitions on demand. For tool-heavy enterprise deployments, this dramatically reduces the token overhead of each request — a meaningful cost savings at scale.
Sources: OpenAI Blog (March 5, 2026), TechCrunch (March 5, 2026), The Verge (March 5, 2026), NxCode (March 2026)