If you rely on AI coding agents for your daily development workflow, you are intimately familiar with the sudden roadblock of hitting a rate limit.
Managing multiple AI subscriptions is expensive, but nothing disrupts a deep programming flow state quite like watching your tool burn through its usage allowance, followed by a polite notification to come back later.
Claude Code is undeniably powerful, but it has a habit of chewing through an almost comical number of tokens. When it navigates a codebase, calls tools, and reasons through problems, it is constantly consuming compute. After hitting my limits a few times too many,
I decided to audit how I was using my subscription. It turns out, optimizing your token runway doesn’t require a complex workaround. I changed a single configuration setting and slashed my output token consumption by nearly half.
The Mechanics of Adaptive Reasoning in Claude
When you hand an LLM a complex prompt, it doesn’t just blindly stream out the first line of code. It engages in internal reasoning, working through logic steps before generating the final output.
This “thinking” phase is what makes newer models so capable, but it is also a massive token sink. You pay for that internal dialogue in both latency and usage limits.
Anthropic built a system around this exact tradeoff called adaptive reasoning. In theory, this allows the model to scale its processing power based on the difficulty of the task.
However, Claude still relies on user guardrails to know exactly how much leash it has to overthink a problem. In Claude Code, you control this through the /effort menu.
The application offers five distinct effort levels. “Low” is heavily restricted for tightly scoped tasks where speed is the priority. “Medium” explicitly reduces token usage for cost-sensitive work. “High” is the default setting for Sonnet, designed to balance intelligence and token spend.
Beyond that, you have “Xhigh” and “Max” for extreme reasoning, though Anthropic warns that pushing it to the maximum often yields diminishing returns. There is also an “Ultracode” toggle that pairs high reasoning with dynamic orchestration.
The problem is that by leaving the environment on its default “High” setting, you are essentially paying a premium token tax for routine, everyday tasks that simply don’t require deep philosophical reasoning to complete.
Testing the Medium Effort Sweet Spot
To see if dropping the effort level would break my workflow, I set up a controlled benchmark. I left Claude Code on its default High setting and fed it five distinct engineering tasks: fixing a known bug, writing a new search feature, refactoring legacy code, expanding test coverage, and tracking down a performance bottleneck.
I logged the usage report for each. Then, I reset the codebase to its exact original state, switched Claude to Medium, and ran the identical prompts.
The reduction in token burn was immediate. On the simple bug fix, Medium output 1,500 tokens compared to High’s 2,000 a solid 25% drop. But as the tasks scaled, the savings became massive.
When instructed to build the search feature, High chewed through 14,400 output tokens. Medium completed the exact same build using only 7,500 tokens, which is a 48% reduction.
Across the board, the data held up. Medium used 54% fewer tokens on the refactoring job, 40% less on test coverage, and 39% less while hunting down the performance bottleneck.
Crucially, the code quality did not degrade. Medium didn’t hallucinate or take shortcuts; it arrived at the exact same technical conclusions as the High setting.
When tasked with finding the performance issue, both effort levels independently isolated the identical under-tested function, identified the same bug, and applied the exact same one-line fix to get all 12 tests passing.
I won’t be abandoning the High effort setting entirely. If I need to architect a complex database migration or debug a convoluted race condition, giving the model more room to reason is a smart investment.
But for standard daily development, Claude simply doesn’t need to overthink. By typing /effort and stepping down to Medium, you can effectively double your usage limits without sacrificing a single drop of code quality.
Source: Official XDA Developers, "I Changed One Setting in Claude Code, and My Token Burn Dropped by 45%"
