Opus 4.7 low thinking surpasses Sonnet 4.6 maximum, Anthropic releases first intelligent gymnastics control tuning guide According to Beating monitoring, Anthropic has released its first official development guide, which provides in-depth information on the resolution limit, depth of thought ratio, and cache cost reduction mechanism of Claude 4.6 and Opus 4.7 in computer and browser control scenarios. The screen resolution directly determines the accuracy of the intelligent agent's clicks. Claude 4.6 has a maximum length of 1568 pixels for analyzing screenshots, while Opus 4.7 has a maximum length of 2576 pixels. Once the screenshot exceeds the limit, the API server will automatically proportionally reduce the image, which can cause misalignment and drift 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 by Claude 4.6) or 1080p (recommended by Opus 4.7) in advance on the client side. Interface manipulation mainly relies on visual perception and element positioning, and does not require high requirements for long chain logical reasoning. Tests have shown that Opus 4.7 can match Sonnet 4.6's maximum thinking depth (max) at low levels of manipulation, and the token cost is only one tenth of the latter. The official suggestion is to set the thinking option to high. Compared to max depth, not only is the token consumption halved, but the success rate is also completely balanced. It is necessary to avoid enabling max to prevent the model from overthinking and causing the bill to double. Due to the fact that a single screenshot can consume up to 1800 tokens in the context, the official has proposed a three-layer cost reduction plan: one system level cache breakpoint is permanently resident, and the other three breakpoints are dynamically allocated to the execution results of the tool in recent rounds; Perform rolling pruning on the client side, retaining only the last 3 screenshots in context and replacing the rest with placeholders; Trigger summary compression when the context depth approaches 90%. In addition, the API introduces the batch tool comput_match, which supports single call packaging and execution of multiple visually independent operations; And provide an intelligent agent advisor mechanism (Advisor Tool), allowing the main model to directly summon high-order Opus models in the background to audit execution steps. Developers can also significantly improve task success rates by recording Teach Mode, which records the user's actual operation trajectory and serves as a reference for instructions during playback. [Original link]