OpenAI announced on 11 March 2026 that its Responses API now includes a built-in computer environment, allowing models to execute code, browse the web, and interact with files without external tooling [source]. The feature, described as moving "from model to agent," gives API users access to a sandboxed Linux container where the model can run Python, install packages, and perform multi-step tasks autonomously.

The announcement states that developers can enable the environment by setting a single parameter in API calls. OpenAI positions the feature as reducing integration overhead for applications requiring code execution or web access, previously handled through third-party frameworks or custom infrastructure.

No independent testing of the environment's behaviour under adversarial prompts or edge cases has been published. The announcement does not specify rate limits, timeout policies, or how the system handles requests that attempt to access restricted resources. OpenAI states the environment is "isolated" but does not detail the sandboxing mechanism or whether output is filtered for sensitive data.

The feature arrives as multiple providers experiment with agentic capabilities. Anthropic's Claude has offered similar tooling since late 2024, and Google's Gemini API includes code execution in select tiers. OpenAI's implementation differs by bundling the environment directly into the API rather than requiring developers to supply execution infrastructure.

The announcement includes sample code and a link to updated API documentation. No pricing changes were disclosed. OpenAI did not publish benchmarks comparing task success rates with and without the environment, nor did it address how the system handles non-deterministic behaviour when models retry failed operations.