Opus 4.7 Low Thoughtfulness Surpasses Sonnet 4.6 Max Value, Anthropic Releases First Smart Robotics Tuning Guide

According to Sentinel Beat monitoring, Anthropic has released its first official Developer Guide, which provides in-depth details about the resolution limits, depth 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 agent's clicks. Claude 4.6 has a maximum screenshot resolution limit of 1568 pixels, while Opus 4.7 is 2576 pixels. Once the screenshot exceeds the limit, the API server will automatically proportionally shrink the image, leading to a misalignment drift between the model-generated click coordinates and the client's original image. Therefore, developers must proactively scale the screenshot to 1280x720 (Claude 4.6 recommended) or 1080p (Opus 4.7 recommended) on the client side.

Interface control mainly relies on visual perception and element positioning, with low requirements for long-chain logical reasoning. Tests have shown that Opus 4.7's control performance at low depth can match Sonnet 4.6's maximum depth, with a token cost only one-tenth of the latter. The official recommendation is to set the thought option to high, which not only halves the token consumption compared to max depth but also maintains an equal success rate. It is advised to avoid enabling max depth to prevent a doubled bill due to model overthinking.

Since a single screenshot consumes up to 1800 tokens in context, the official document provides a three-tier cost reduction plan: establish a system-level cache breakpoint and dynamically allocate three breakpoints to the execution results of the most recent rounds; perform scroll pruning on the client side, retaining only the last 3 screenshots in context, and replacing the rest with placeholders; trigger summary compression as the context depth approaches 90%.

Additionally, the API has introduced the batch tool computer_batch, which supports packing multiple visual-independent operations for execution in a single call; and has implemented an Agent Advisor mechanism, allowing the main model to directly summon higher-order Opus models in the background to audit the execution steps. Developers can also significantly improve task success rates through the Teach Mode, which records the user's actual operating trajectory for playback as an instruction reference.