GPT-5.4 Pro

GPT-5.4 Pro is OpenAI's most capable model, the flagship of the GPT-5.4 family that unifies frontier reasoning, coding, and computer use into a single system. It features a 1-million-token context window (272K standard, 1M in Codex experimental mode), 128K max output tokens for generating entire codebases in one pass, native computer use that interacts with desktops through screenshots, controls mouse and keyboard, and writes Playwright code for browser automation, and tool search — an agentic capability that loads tool definitions on demand instead of all at once, saving tens of thousands of tokens per request.

On professional benchmarks, GPT-5.4 Pro tops the GDPval-AA leaderboard at 1667 Elo, ahead of Claude Sonnet 4.6 at 1633 and Opus 4.6 at 1606. It achieves 87.3% on spreadsheet modeling (vs 68.4% for GPT-5.2), and human raters preferred GPT-5.4 presentations 68% of the time over GPT-5.2. On Humanity's Last Exam, it scored 52.1%, breaking the 50% threshold for the first time among any model. FrontierMath reached 47.6% (vs 40.3% for GPT-5.2), ARC-AGI-1 hit 93.7%, and ARC-AGI-2 reached 83.3% in Pro mode (vs 73.3% for standard GPT-5.4).

The computer use and agentic benchmarks are where GPT-5.4 Pro truly stands apart. On OSWorld-Verified, it scores 75.0% — exceeding human performance of 72.4%, with the previous top model being Kimi K2.5 at 63.3%. This is the first time any AI model has surpassed human-level performance on this benchmark. Web browsing capabilities are equally impressive: 89.3% on BrowseComp (Pro variant; standard GPT-5.4 at 82.7%), 67.3% on WebArena-Verified, and 92.8% on Online-Mind2Web. Toolathlon scored 54.6%, demonstrating strong autonomous tool use.

The model delivers a 33% reduction in false claims and 18% fewer responses containing any errors compared to predecessors. Its steerability is significantly enhanced — it outlines a plan before continuing and allows mid-response adjustments, giving users more control over the output direction. Tool search saves tens of thousands of tokens per request by loading tool definitions on demand rather than including all definitions in every context window.

Native computer use operates via screenshots, mouse and keyboard control, and Playwright browser automation, enabling the model to interact with desktop software, fill out web forms, navigate multi-tab workflows, and execute complex GUI-based tasks autonomously. This makes GPT-5.4 Pro uniquely suited for enterprise automation scenarios that require interacting with legacy software, web applications, and desktop tools.

Safety evaluations include a chain-of-thought controllability study showing that models cannot effectively hide their reasoning, with controllability rates between 0.1% and 15.4%. OpenAI expanded its cyber safety stack and reduced refusals compared to GPT-5.2, making the model more helpful on legitimate edge-case queries without compromising on genuinely harmful requests.

At $30/M input and $180/M output, GPT-5.4 Pro is priced for teams that need the absolute best performance on complex reasoning, long-horizon agentic workflows, professional-grade analysis, and tasks where exceeding human-level performance on computer use and web browsing is critical. For most production workloads, the standard GPT-5.4 at $2.50/$15 offers the same architecture at a fraction of the cost.