Anthropic moves Computer Use out of beta, ships native sandbox primitive
Claude's screen-grounded agent loop graduates with new tool-use primitives, an isolated sandbox, and tighter rate-limit policy for production deployments.
Anthropic moved its Computer Use capability into general availability today, exiting a six-month beta that had been gated behind a developer waitlist. The release adds a hosted sandbox that isolates browser and shell sessions per agent run, plus first-class tool primitives for keyboard, mouse, and clipboard actions.
What changed. Computer Use is GA. There is now a native isolated sandbox, deterministic screenshot sampling, and a published rate-limit policy for production traffic. Pricing for screenshot tokens is unchanged, but session-based billing replaces per-action billing.
Why it matters. This closes the largest operational gap between research demos and production deployments — sandbox lifecycle and screenshot cost predictability. Teams that had built their own VM-per-task harnesses can now deprecate that scaffolding.
Builder takeaway. If you are running browser agents on EC2 or Fly machines, audit the session pricing against your current per-task cost. Expect a 20-40% reduction for short-horizon tasks (under 90 seconds) and a small increase for long-horizon research tasks.
The company also published an evaluation harness for screen-grounded tasks that aligns with WebArena and OSWorld scoring, removing one of the more visible inconsistencies between vendors.