OpenAI computer-use-preview

computer-use-preview arrived on 11 March 2025 as part of the largest single release OpenAI had made for developers: the Responses API, the Agents SDK and three built-in tools — web search, file search and computer use. This model was the engine of the last of those. It took screenshots as input and returned the actions to perform on them: click here, type this, scroll there. Its most surprising number was the context window. At 8,192 tokens it was one-sixteenth of what GPT-4o offered at the same time, with a ceiling of 1,024 output tokens. A model built to look at pictures of screens was given less room than a model built to read documents, which forced developers into short, tightly scoped agent loops rather than long sessions. Input was text and images, output text only, with reasoning token support and a knowledge cutoff of October 2023. It ran in the Responses API and Batch only — never in Chat Completions. The price was 3 dollars per million input tokens and 12 per million output, plus a separate fee for every tool call. OpenAI deprecated it on 22 April 2026 with the short notice reserved for preview models and shut it down on 23 July 2026. By then the capability had not disappeared but moved: GPT-5.4 gained computer use as a built-in tool of the Responses API, so the job that once needed a dedicated model became a parameter of a general one.

▸Videos

No videos yet.

▸Reviews