OpenAI Launches GPT-5.4 with Built-In Computer Vision and Desktop Control

Just two days after shipping GPT-5.3 Instant, OpenAI has unveiled its next-generation models — GPT-5.4 and GPT-5.4 Pro. The headline feature: native computer vision that lets the model see a screen and control a desktop using mouse and keyboard.

"GPT-5.4 Thinking and GPT-5.4 Pro are rolling out now in ChatGPT. GPT-5.4 is also now available in the API and Codex. GPT-5.4 brings our advances in reasoning, coding, and agentic workflows into one frontier model." — OpenAI (@OpenAI), original post

Availability and Pricing

The standard GPT-5.4 is available through ChatGPT's web interface, the API, and Codex. GPT-5.4 Thinking is open to Plus, Team, and Pro subscribers. GPT-5.4 Pro targets Pro-tier users and Enterprise clients, with API access included.

Base model pricing sits at $2.5 per 1M input tokens and $15 per 1M output tokens. The Pro version costs considerably more at $30 and $180 per 1M tokens, respectively.

Why This Matters

GPT-5.4 is OpenAI's first model with built-in computer vision and full desktop control. It can operate a mouse and keyboard based on screenshots and automate workflows through Playwright. Risk tolerance is configurable per use case.

On the OSWorld-Verified benchmark for desktop management, GPT-5.4 completed 75% of tasks — surpassing the previous version (47.3%) and human performance (72.4%). Enhanced visual perception shows across other tests too: MMMU-Pro (understanding and reasoning) hit 81.2% vs. 79.5% for GPT-5.2, while OmniDocBench (document analysis) error rates dropped from 0.140 to 0.109.

Professional Task Performance

On GDPval, a benchmark evaluating task completion across 44 professions, GPT-5.4 scored 83% — up from GPT-5.2's 70.9%. The model now performs at or above the level of domain specialists.

Developers focused heavily on spreadsheets, presentations, and documents. On tasks at the level of a junior investment banking analyst, GPT-5.4 scored 87.3% compared to 68.4% for GPT-5.2. Evaluators preferred the new model's presentations in 68% of cases, citing better aesthetics, variety, and effective use of image generation.

Factual accuracy also improved significantly. When tested on prompts containing deliberate errors, individual false claims appeared 33% less often, and full responses contained errors 18% less often than GPT-5.2.

Coding and Tools

In programming, GPT-5.4 matches the specialized GPT-5.3-Codex while running faster. Codex now includes a /fast mode that speeds up code generation by 1.5x with no quality loss. Internal tests showed strong results on complex front-end development tasks.

An experimental Playwright (Interactive) skill lets the model visually debug web and Electron applications, testing its own code in real time during the writing process.

Another addition is Tool Search. Previously, the system had to preload descriptions of all available plugins into context, adding thousands of unnecessary tokens per request. Now the model receives only a base list and dynamically loads the parameters it needs. In tests using MCP Atlas, token consumption dropped by 47% with no accuracy loss.

Web search performance also improved: BrowseComp scores rose by 17%, with the Pro version reaching a record 89.3%.

Steerability and Context

GPT-5.4 Thinking in ChatGPT now displays an action plan before tackling complex queries. Users can adjust the direction mid-stream without restarting generation. The feature is live on the website and the Android app, with iOS support coming soon.

The model also maintains context more effectively in extended conversations and dedicates more reasoning time to complex tasks, keeping responses coherent and relevant when handling large volumes of information.