Anthropic has published its first official developer guide for building agents that control computers and browsers, spilling the beans on resolution caps, thinking depth ratios, and caching cost reductions for Claude 4.6 and Opus 4.7.
Screen resolution directly determines how accurately an agent can click. Claude 4.6 has a long-edge limit of 1,568 pixels for screenshots, while Opus 4.7 tops out at 2,576. If your screenshot exceeds those limits, the API server will automatically downscale the image proportionally — causing the model’s click coordinates to drift from your client-side original. Developers need to pre-scale screenshots to 1280x720 (recommended for Claude 4.6) or 1080p (recommended for Opus 4.7) on the client side.
Interface control relies mostly on visual perception and element location, not long chains of logical reasoning. Testing shows Opus 4.7 at low thinking depth matches Sonnet 4.6 at max thinking depth — and costs just one-tenth the tokens. Anthropic recommends setting thinking to "high"; compared to "max", token consumption is halved while success rates stay identical. Avoid "max" to prevent the model from overthinking and doubling your bill.
Each screenshot can consume up to 1,800 tokens in the context window. Anthropic suggests a three-tier cost-reduction plan: keep one system-level cache breakpoint, dynamically assign three breakpoints to recent tool execution results; perform scroll pruning on the client to keep only the last three screenshots in context, substituting the rest with placeholders; and trigger summary compression when context depth approaches 90%.
The API also introduces a batch tool called computer_batch, which lets you execute multiple vision-independent operations in a single call. There’s also an Advisor Tool that allows the main model to summon a high-end Opus model in the background to audit execution steps. Developers can further boost task success rates with Teach Mode — recording real user actions and replaying them as instruction references.