GPT-5.4 lands with 1M context and native computer control
OpenAI's GPT-5.4 hits 92.8% on web browsing benchmarks and becomes the first AI to beat humans on OSWorld desktop tasks.
GPT-5.4 dropped on March 5, rolled out simultaneously across ChatGPT, the API, and Codex. The headline feature: a 1 million token context window and first-party computer-use — screenshot-only browser automation that actually works at scale.
On benchmarks, it scores 92.8% on Online-Mind2Web (vs ChatGPT Atlas's 70.9%) and 75.0% on OSWorld, edging past the 72.4% human expert baseline — the first general-purpose model to clear that bar. MMLU sits at 88.5%, fractionally ahead of Claude 4's 87.9%.
API access is live today. The 1M context window is an opt-in experimental feature requiring explicit config — without it, you're on the standard 272K window.
Why it matters: computer use at 92%+ web accuracy isn't a demo anymore — it's infrastructure for autonomous agents that can actually browse and act.
Sources
Independent coverage
- GPT-5.4 Complete Guide: Features, Benchmarks, Pricing almcorp.com
- GPT-5.4 Review: Features, Benchmarks & Access (2026) buildfastwithai.com
Written by an AI pipeline from the sources above. Methodology · Report an error
Feed, daily deep-dive and bytes — readable offline, with push alerts for the topics you follow.