OpenAI Agents API adds hosted browser and sandboxes
OpenAI Agents API update introduces hosted browser automation, persistent sandboxes, and GPT-6.1 Sol support. Service remains in public beta.
The weekly update delivers cloud-hosted web browsing, reusable virtual environments, and GPT-6.1 Sol support, while pricing details remain undisclosed.
OpenAI rolled out a major update to the OpenAI Agents API, targeting engineers building autonomous software workflows. Announced by the official @OpenAIDevs account and detailed by engineer Steve Coffey on October 3, 2026, the changes address two major friction points: browser automation without external infrastructure, and the routine disposal of virtual execution environments between runs. The package also adds native access to the GPT-6.1 Sol model and an updated console view for tracking subagent handoffs.
What changes in the OpenAI Agents API: cloud browser
The headline addition, dubbed “Computer use,” runs browser automation directly on OpenAI servers. Instead of provisioning Selenium clusters or managing headless Playwright containers, client applications can trigger web navigation, form submissions, and data extraction through standard API calls.
To prevent credential leakage and untrusted navigation, OpenAI built in an origin restriction. Developers must whitelist target domains through the browser_origin_access parameter. Login credentials and sensitive tokens stay with the client application, allowing teams to isolate what the agent can touch.
Persistent sandboxes and machine tiers
Until now, developers faced heavy cold-start delays when orchestrating code-execution sandboxes. Each session launched clean, forcing scripts to reinstall Python packages, pull dependencies, and reconfigure disk state from scratch. The updated API allows teams to reuse active sandboxes across subsequent calls, keeping installed libraries and local files intact.
OpenAI structured these virtual environments into two capacity tiers: “light” and “beefy.” While the setup supports everything from quick script execution to heavy data processing, the company has not published hourly or per-minute pricing for the heavier tier.
GPT-6.1 Sol and self-reported performance
Alongside infrastructure changes, the release connects natively to GPT-6.1 Sol, a model built for complex code generation and parallel multi-agent coordination. Inside the developer dashboard, an updated observability panel maps out how a primary agent breaks down and delegates work to subagents.
According to technical figures published by OpenAI engineers, the release achieved a 99.97% per-turn reliability rate, fewer dropped Server-Sent Events (SSE) streaming connections, and a 20% latency cut for tool calls. These statistics stem strictly from OpenAI internal benchmarks and have not been independently tested or audited by third parties.
Infrastructure competition and beta limits
By bundling managed browsers and execution sandboxes, OpenAI enters territory long held by specialized providers like E2B, Modal, and Daytona. Consolidating these services reduces setup time and brings computing, browsing, and reasoning onto a single bill. It also ties teams more firmly to OpenAI’s proprietary ecosystem.
The Agents API remains in public beta under the OpenAI-Beta: agents=v1 header, with no target date for general availability. Furthermore, rumors regarding direct parity or integration with AWS Bedrock remain unconfirmed speculation, as Amazon has made no announcement regarding the tool.