GPT-5.4 Sets New Records for Computer Use — AI That Operates Your PC

On March 5, 2026, OpenAI shipped GPT-5.4 — a focused update to GPT-5 that pushed computer-use capabilities far beyond any previous model.

Key benchmark results at launch:

• OSWorld-Verified: Record score — surpassing all prior models including Claude Sonnet 4.5
• WebArena Verified: Record score on automated web navigation tasks
• GDPval: 83% on OpenAI's internal test for knowledge work tasks — the first time any model crossed 80%
• Context window: Expanded to 272,000 tokens from GPT-5's original 128K

GPT-5.4 can operate a computer the way a human does — clicking buttons, navigating browsers, reading screens, filling out forms, and executing multi-step workflows across applications. It doesn't just generate code; it can run the code, see the result on screen, and fix errors in a loop.

OpenAI also shipped GPT-5.5 Instant alongside it — a 400K token context window model with reasoning capabilities priced for high-volume API workloads.

Computer use is the threshold where AI transitions from answering questions to getting things done on your behalf.

The implications of an 83% GDPval score are significant. GDPval tests whether a model can complete real knowledge work tasks: researching topics, synthesizing documents, drafting reports, filling spreadsheets, sending structured outputs to systems. At 83%, GPT-5.4 can reliably complete most office tasks without human intervention.

This is the beginning of a genuine agentic economy:

• Customer support workflows that don't require a human in the loop
• Data entry and form processing that runs overnight autonomously
• Software QA agents that can actually use the UI, not just call APIs
• Personal productivity agents that manage calendars, email, and documents

The model has limitations — it still makes errors on complex visual tasks and struggles with CAPTCHAs by design — but the trajectory is clear.