Opus 4.7 outperforms Sonnet 4.6's maximum low-level thinking; Anthropic releases its first guide to optimizing agent manipulation.

According to Beating's monitoring, Anthropic has released its first official development guide, providing in-depth details on the resolution limits, thought depth ratios, and caching cost reduction mechanisms of Claude 4.6 and Opus 4.7 in computer and browser control scenarios. Screen resolution directly determines the accuracy of agent clicks. Claude 4.6 has a maximum long side limit of 1568 pixels for parsing screenshots, while Opus 4.7 has a limit of 2576 pixels. Once the screenshot exceeds the limit, the API server will automatically scale the image proportionally, which will cause a misalignment between the click coordinates generated by the model and the original image on the client side. Therefore, developers must scale the screenshot to 1280x720 (recommended for Claude 4.6) or 1080p (recommended for Opus 4.7) on the client side beforehand. Interface control mainly relies on visual perception and element positioning, and does not require high levels of long-chain logical reasoning. Tests show that Opus 4.7's performance at low thinking depth (low) matches Sonnet 4.6's maximum thinking depth (max), with token costs only one-tenth of the latter. The official recommendation is to set the thinking option to high; compared to max depth, not only is token consumption halved, but the success rate is also completely equal. Max depth should be avoided to prevent the model from overthinking and doubling the cost. Since a single screenshot consumes a maximum of 1800 tokens in the context, the official solution provides a three-layer cost reduction approach: maintaining one system-level cache breakpoint and dynamically assigning three other breakpoints to the execution results of the most recent tool rounds; performing rolling pruning on the client side, retaining only the three most recent screenshots in the context, replacing the rest with placeholders; and triggering summary compression when the context depth approaches 90%. Furthermore, the API introduces the batch tool `computer_batch`, supporting the single-call execution of multiple operations without visual dependencies; and provides an agent advisor mechanism (Advisor Tool), allowing the main model to directly invoke a higher-order Opus model in the background to audit the execution steps. Developers can also significantly improve task success rates by recording a guided mode (which records the user's actual actions and uses them as a reference for instructions during playback).