OpenAI's GPT-5.4 Adds Native Computer Use — Scores 75% on OSWorld, Beating Average Humans

OpenAI launched GPT-5.4 on March 5, combining reasoning, coding, and agentic workflows into a single model — and the most consequential addition is native computer use.

Previous OpenAI models could write code or plan tasks. GPT-5.4 can actually execute them. In Codex and the API, the model can interpret screens, control a mouse and keyboard, and operate across applications to complete complex workflows. This isn’t a plugin or a wrapper — it’s built into the model’s architecture.

The benchmark that matters

On OSWorld-Verified, the standard benchmark for computer-use agents, GPT-5.4 scored 75%. The average human performance on the same tasks is 72.4%. That puts GPT-5.4 above the average human baseline for desktop task automation — the first OpenAI model to cross that line.

On GDPval, which measures performance on professional-grade tasks, GPT-5.4 matches or exceeds industry professionals in 83% of comparisons.

What’s actually in the release

Three model variants shipped: GPT-5.4 Standard, GPT-5.4 Thinking (reasoning-first, optimized for multi-step problems), and GPT-5.4 Pro (maximum capability, higher cost). The context window across all variants supports up to 1 million tokens — the largest OpenAI has made commercially available.

Tool Search is new and practical: when an agent operates in a large ecosystem of hundreds of connectors and functions, finding the right tool at inference time is itself a non-trivial problem. GPT-5.4 addresses this with built-in tool search, allowing agents to dynamically select from large tool sets without the developer pre-specifying every possible path.

Skills — reusable, shareable workflow bundles that include instructions, examples, and code — are available in ChatGPT. They can be applied automatically when the model identifies a relevant context.

The agentic shift

The pattern across all these features is the same: GPT-5.4 is designed to complete tasks, not assist with them. Computer use means the model can act on a computer. Tool search means it can navigate large systems. A 1M token context means it can hold an entire project in working memory. Thinking variants mean it can plan before it acts.

OpenAI’s competitive position in the agentic race — against Anthropic’s operator-based agent framework, Google’s Gemini agents, and open-source alternatives — depends on this model performing in production the way it does in benchmarks.

That’s always the question with computer-use systems. OSWorld-Verified is a controlled environment. Real enterprise desktops are not.