GPT-5.4: Agentic Computer Use Is the Product, Not Chat Elo
OpenAI released GPT-5.4 on March 5 with native computer use, tool search and up to 1M context. It reported 75.0% on OSWorld-Verified against a 72.4% human baseline; these are vendor research results, not a general productivity guarantee.
Correction / update, October 14, 2026: Removed unverified historical Arena scores and later mini/nano launch context. The computer-use result is OpenAI’s OSWorld-Verified evaluation.
OpenAI positioned GPT-5.4 around agentic computer use, configurable reasoning effort, tool search and long context. Its release evaluations do not establish a historical Arena preference ranking.
OpenAI unveiled GPT-5.4 (March 5) with Pro as its most capable model family for agent-style professional work: native computer-use APIs; configurable reasoning effort; context up to 1M tokens; coding/reasoning gains; Tool Search for dynamic tool calling; ChatGPT + API + Codex. Company materials: ~75% on OSWorld (slightly above human baseline 72.4%); strong SWE-bench-class coding. Resource shift toward developer and agent tools.
Claims vs checks#
Launch features are OpenAI primary. OSWorld/SWE figures are vendor (effort settings matter).
Limits#
- Computer-use success rates transfer unevenly outside eval harnesses.
- 1M context availability/pricing varies by SKU.
- Codex 1M-context support was experimental at launch.
Sources#
- OpenAI: “Introducing GPT-5.4” (March 5, 2026).