OpenAI Adds Hosted Browser Control to the Agents API
The September 29 update lets applications run agent tasks in an OpenAI-hosted browser while keeping website access and sign-in decisions in the app.
Edited by Tyronne Panaino
OpenAI added computer use to the Agents API on September 29, 2026, according to the official API changelog. The update lets an agent complete tasks in an OpenAI-hosted browser while the developer's application handles website access approvals and sign-in.
The change matters to developers already evaluating the Agents API because it expands the managed service beyond sessions, tools and sandboxes into browser interaction. It does not remove the application from the permission boundary: the brief release record explicitly leaves access decisions and authentication handling with the integrating software.
The delta from the September beta
OpenAI's changelog records the Agents API public beta on September 10. At that point, the company described a managed Codex harness in which OpenAI handled session orchestration, context compaction and recovery. The same release record listed durable sessions, progress streaming, external tools, MCP servers and a choice between OpenAI-hosted or connected sandboxes.
The September 29 entry adds computer use to that existing service. The supported conclusion is narrow but material: an Agents API workflow can now include work inside a browser hosted by OpenAI. This is separate from releasing a new model and separate from merely connecting a conventional API tool. The browser becomes another environment in which the agent can act.
Applications still own the access boundary
The changelog says website access approvals and sign-in are handled by the application. For an engineering team, that means the hosted browser does not make identity and authorization design disappear. The application still has to decide when access is acceptable and how a sign-in step fits into its own user experience.
That responsibility split is more important than a generic claim that an agent can browse. Browser tasks can reach changing interfaces and authenticated services, so teams evaluating the feature will need to test how their approval flow behaves and how completed work is checked. Those are implementation questions for each application, not outcomes established by the announcement.
Evidence quality and open questions
The fetched evidence is one first-party changelog. It establishes the release date, hosted-browser scope and the stated responsibility for approvals and sign-in. It does not provide independent tests of task completion, reliability, latency or security outcomes. It also does not specify pricing, supported websites, geographic availability or a production service-level commitment for this feature.
The next verifiable checkpoint is fuller product documentation or independent testing that shows how the browser feature behaves across real workflows. Until then, the reliable claim is that computer use has joined the Agents API, not that it will complete every browser task safely or successfully.
Status
Confirmed. OpenAI's API changelog records the September 29 feature addition. Internal confidence is medium because the evidence is first-party and this run found no independent operational evaluation.
Sources
Update note: Last reviewed 2026-09-30. We will revise this post if OpenAI publishes availability, pricing, service-level or independent evaluation details.
Sources
- OpenAI API — Changelog — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.