Claude Code Effort Levels Cut Token Usage by 45%
A new setting allows users to limit AI reasoning for simple tasks, significantly reducing token consumption without switching models.
Key points
- Claude Code now offers five effort levels, from Low to Max, to control reasoning depth.
- Setting the effort to Medium instead of High reduced token usage by 45 percent in tests.
- The High level is the default, but Max allows unrestricted token spending for complex problems.
Users of AI coding assistants often face frustration when usage limits interrupt their workflow. A significant portion of this issue stems from the tools themselves expending excessive computational resources on routine tasks. The latest update to Claude Code addresses this inefficiency by introducing granular control over the model's reasoning effort, allowing developers to match the tool's intensity to the complexity of the job.
Previously, developers had to manually switch between different AI models to manage costs, opting for faster, cheaper options for simple edits and powerful, expensive models for complex debugging. The new feature eliminates this friction by letting users adjust the effort level within a single model. This adjustment trades some depth of analysis for greater token efficiency, ensuring that simple requests like renaming variables do not consume the same resources as architectural planning.
Five levels define reasoning depth
The system offers five distinct settings: Low, Medium, High, Xhigh, and Max. Each level dictates how much room the model has to reason through a problem. The High setting is currently the default for the Sonnet 5 model, aiming for a balance between intelligence and consumption. Lowering the setting encourages the AI to be more economical, while raising it provides more space for complex logic.
The highest setting, Max, removes most restraints on token spending to achieve the highest possible capability. While effective for difficult problems, leaving this enabled for all tasks is inefficient. It is designed for scenarios requiring deep analysis, not for minor formatting changes or typo corrections where the extra reasoning adds cost without significant benefit.
Testing reveals significant cost savings
Practical testing by XDA Developers highlighted the tangible benefits of adjusting these settings. By dropping the effort level from High to Medium for routine tasks, token usage was reduced by 45 percent. This demonstrates that a substantial amount of previous consumption was unnecessary overhead for simple coding tasks.
The trade-off is clear: lower effort levels may miss subtle nuances in complex code, while higher levels waste resources on simple edits. Users must now make a conscious decision about the required depth for each request. This shift places the responsibility on the developer to calibrate the tool, rather than relying on a one-size-fits-all default that often overspends.
Efficiency requires manual calibration
Adopting this feature requires a change in habit. Instead of treating the AI as a static tool, users must actively manage its reasoning depth. This approach aligns the tool's behavior with human cognitive patterns, where simple tasks receive quick responses and complex problems receive sustained focus.
The catch is that users must accurately judge the complexity of their own tasks. Misjudging a simple task as complex leads to wasted tokens, while misjudging a complex task as simple may result in suboptimal code. However, for the majority of daily coding activities, the default high-effort setting is likely overkill, and adjusting down offers a practical path to better resource management.
