Some developers are using the Claude Code harness to access cheaper non-Anthropic models, including GPT-5.6 Sol, via proxies and services like OpenRouter

You can generate 2x more code output from AI assistants by letting them handle execution.

Anthropic has analyzed 400,000 sessions using its Claude Code tool to understand user interaction patterns. The research reveals that users who delegate execution decisions to Claude Code, rather than micromanaging every step, achieve better results. In these effective sessions, users typically define the desired outcome and constraints, while Claude handles the implementation details, performing an average of 10 to 16 actions per turn compared to fewer actions when users retain more control. The analysis also highlights that user expertise remains crucial; individuals with sufficient domain knowledge can better frame instructions, identify errors, and guide Claude more effectively, leading to significantly more actions and output per prompt. Finally, the study suggests that users should provide Claude Code with specific methods to verify its own work, moving beyond generic requests to targeted checks that anticipate potential failure points.

The analysis indicates a significant productivity gain for AI coding assistants when users shift from detailed instruction to outcome-based delegation. By defining goals and constraints and allowing the AI to manage the 'how,' users can achieve twice the number of actions per interaction, demonstrating that AI's efficacy is directly tied to the autonomy granted within defined boundaries. This suggests a paradigm shift in how developers should approach AI collaboration, prioritizing strategic direction over tactical micromanagement.

User expertise plays a critical role, not in dictating code, but in effective validation and problem framing. Domain knowledge allows users to recognize successful outcomes and guide the AI past errors, amplifying its output and efficiency. The research suggests that providing specific, verifiable checks for AI-generated work, rather than generic assurances, is key to unlocking deeper capabilities and ensuring quality, impacting how developers interact with and trust AI coding partners.