By now, you’ve got doubtless had the belief that AI instruments are costly. You may’ve realized that by trying on the worth of your rising pile of month-to-month subscriptions. Or the thought may’ve hit you once you got the polite “you’ve hit the limit! please come back later” message after utilizing your favourite AI software for what felt like 5 minutes.
I’ve skilled each, however Claude Code has made me significantly acutely aware of the latter. Coding brokers can burn by an virtually comical variety of tokens whereas they’re poking round a codebase, reasoning by an issue, calling instruments, and sometimes doing much more considering than the duty appears to warrant. So naturally, I’ve began in search of methods to get essentially the most out of Claude Code subscription. Seems, altering one setting is all it took to make a surprisingly large distinction.
Claude Code helps you to resolve how onerous it ought to assume
Extra considering is not at all times higher
While you ask an LLM a query, it would not immediately spit out a solution. As an alternative, it really works by the issue internally earlier than producing the response you really see. The extra advanced the duty, the extra reasoning it might have to do. After all, further considering means extra tokens being consumed alongside the best way. This interprets to each extra ready in your finish and extra of your utilization allowance being burned by within the background. That tradeoff could be value it for genuinely troublesome duties, but it surely makes so much much less sense when the job is comparatively simple.
Claude Code already has a system designed round that precise tradeoff. Its newer fashions use what Anthropic calls adaptive reasoning, which suggests Claude can resolve whether or not a specific step wants deeper reasoning and the way a lot considering is definitely worthwhile. In concept, that lets routine duties transfer alongside rapidly whereas extra difficult issues get the additional consideration they want.
Claude handles a few of that routinely, however you continue to get a say in how a lot room it has to assume. The primary manner to do this is thru Claude Code’s effort setting. Every effort stage trades token utilization in opposition to functionality, so decreasing it encourages Claude to be extra economical with its reasoning, whereas elevating it offers the mannequin extra room to work by troublesome issues.
There are 5 precise effort ranges on the latest supported fashions. Low is meant for brief, tightly scoped duties the place velocity issues greater than most intelligence. Medium reduces token utilization for cost-sensitive work, with the opportunity of sacrificing some functionality in return. Excessive is designed to steadiness token utilization and intelligence, and is the default on most supported fashions. Xhigh offers Claude extra room for deeper reasoning at a better token price, whereas Max pushes that even additional. Anthropic even warns that Max can hit diminishing returns and turn into susceptible to overthinking.
Claude Code additionally presents Ultracode in the identical /effort menu, although technically it is not one other mannequin effort stage. As an alternative, it combines Xhigh reasoning with Claude Code’s dynamic workflow orchestration for extra substantial duties.
All it took was dropping Claude Code to Medium
Claude would not have to overthink every thing
For my testing, I did not do something significantly dramatic. I left Claude Code on Excessive, its default effort stage for Sonnet 5, gave it 5 completely different coding duties, and recorded the utilization report after each. Then I reset the codebase, switched Claude Code to Medium, and gave it the very same prompts once more.
The duties ranged from fixing a bug and including a search characteristic to refactoring code, bettering take a look at protection, and monitoring down a efficiency bottleneck. I additionally made positive each effort ranges began from the identical untouched model of the mission every time, so the one significant variable I used to be altering was the trouble setting.
The distinction confirmed up virtually instantly. On the straightforward bug-fixing job, Medium generated round 1,500 output tokens in contrast with 2,000 on Excessive, a 25% discount. The hole turned a lot bigger once I requested Claude Code to construct a search characteristic: Excessive burned by roughly 14,400 output tokens, whereas Medium completed the identical job with round 7,500 — almost 48% fewer.
The sample continued throughout the remainder of my exams. Medium used about 54% fewer output tokens on the refactoring job, 40% fewer whereas bettering take a look at protection, and 39% fewer whereas discovering and fixing a real efficiency drawback.
Add every thing collectively, and Excessive generated roughly 26,000 output tokens throughout my 5 exams. Medium used round 14,300. That is a forty five% drop in output token utilization from altering a single setting.
Extra importantly, Medium did not all of a sudden begin falling aside. It efficiently accomplished all 5 duties, and on a few of the extra open-ended exams, it reached virtually precisely the identical conclusions as Excessive. After I requested each to search out an under-tested piece of logic, for instance, each independently selected the identical task-completion perform, found the identical bug, utilized the identical one-line repair, and completed with all 12 exams passing. After I requested them to hunt for a efficiency difficulty, each noticed the identical precise drawback!
The important thing is choosing the right effort for the job
In the end, I am not suggesting you need to completely abandon Excessive effort and power Claude Code to do every thing on Medium. There are nonetheless going to be difficult debugging classes, architectural choices, and genuinely troublesome issues the place giving Claude extra room to purpose is sensible.
However for the sort of on a regular basis coding duties I examined, Excessive felt more and more troublesome to justify. Medium accomplished the identical 5 jobs whereas producing 45% fewer output tokens, spending much less time processing them, and displaying no apparent drop within the high quality of the outcomes. At this level, I might slightly begin on Medium and solely flip the trouble up when a job really offers me a purpose to.
