How-To Geek's Claude Effort Test Found a 1%-Versus-8% Gap
The test was deliberately modest. The author, a Claude Pro subscriber since launch, picked three tasks they could check themselves: debugging a Linux script, finding time-zone conflicts between meetings, and splitting expenses for a trip. They ran all three in one chat at Low effort, then opened a new chat and ran the same prompts at High effort on the same model.
Both runs produced correct answers, and the presentation was nearly identical. The Low-effort session moved the plan's usage meter by 1%. The High-effort session moved it by 8%. From that the author concluded that everyday knowledge work should default to Low, because Claude "will happily" spend extra tokens even when it can get the right answer with far fewer.
The author also stated the test's limits. They said it does not show that more effort never helps, and that proving that wasn't the goal. The comparison rests on usage-meter percentages shown in screenshots, not on token counts, and each configuration was run once. Because the meter is rounded, "1%" could stand for a range of actual consumption, and the true ratio could sit well above or below 8x.
The author then offered a second tip: when Claude gets stuck, switch to a more capable model instead of raising effort. Their ladder runs from Sonnet to Opus to Fable, the top tier of Anthropic's current lineup. They argue this is more likely to solve the problem and "probably more economical." That second claim comes with no cost or quality measurements, so it's the author's opinion from experience, not a tested finding. The author does allow one case for higher effort: hands-off workflows where you point the model at a task with little prompting and let it brute-force the problem. That trades tokens for your own time.
What the Claude Effort Selector Actually Controls in claude.ai and Claude Desktop
The setting is easy to find, which the author jokes has made it "hidden in plain sight." Anthropic's Help Center says to click the model name next to the send button and choose a level. Per the Help Center, the effort selector is available for Claude Opus 5.5, Fable 5.1, Opus 5, Sonnet 5, Fable 5, Opus 4.7, Opus 4.6, and Sonnet 4.6. Older models in the picker won't show the control.
Anthropic describes the mechanism plainly: the effort level controls how much thinking Claude applies to a response. Higher effort means more thorough responses, but they take longer and use more tokens, so you'll reach your usage limits faster. The menu marks each model's recommended level as "Default," and the Help Center's guidance on that menu matches the How-To Geek advice closely: Low and Medium work well for routine tasks and stretch your usage further. High offers the best overall balance of quality and speed. Above High, extra high (xhigh) is designed for long-running coding and agentic tasks, offering deeper reasoning than high without the full token cost of max.
The feature is recent in the consumer apps. Anthropic added effort control to claude.ai and Cowork with the Claude Opus 4.8 launch on May 28, 2026, as a new control next to the model selector, available on all plans. In the launch post, Anthropic said higher settings make Claude think more often and more deeply to improve responses, while lower settings make it respond faster and use up rate limits more slowly.
The effort selector is separate from the Thinking toggle that sits under it in the same menu. HSKY Lab's tutorial explains that effort sets how thorough Claude is overall, and at low it may skip thinking entirely on a simple problem. It also notes that you can combine Thinking on with Low effort, and concludes that effort is the bigger lever for cost; thinking is mostly about visibility and depth.
That behavior helps explain the How-To Geek result. Anthropic's developer guidance for adaptive-thinking models says the amount of reasoning depends on both the effort setting and how complex the query is. Simple prompts may get a direct answer, higher effort draws out more thinking, and lowering effort can cut thinking tokens when a model overthinks. Time-zone arithmetic and expense splitting are exactly the kind of well-defined tasks where High effort spends tokens that don't change the answer.
Anthropic's Changing Claude Defaults Show the Effort Trade-Off Is Real
The How-To Geek piece says Anthropic recommended Medium as the default for Sonnet and High for Opus when it was written. Anthropic's current records support a narrower point: defaults vary by model and product, and Anthropic has changed them more than once.
For Opus 4.8, Anthropic said the model defaults to High, which it called the best overall balance of quality and user experience. It said "extra" (called "xhigh" in Claude Code) and "max" spend more tokens for better results and recommended "extra" for difficult tasks and long-running asynchronous workflows. On the API side, the platform documentation says Claude Opus 5.5 supports all five effort levels, and medium is the default (Claude Opus 5 and earlier Opus models default to high, so a request that omits effort runs one level lower than it did on Claude Opus 5). For Fable 5, the docs advise starting at High and stepping down to medium or low for routine or latency-sensitive work once your evals show quality holds.
The clearest evidence of the trade-off is Anthropic's April 23, 2026 postmortem on Claude Code quality complaints. On March 4, Anthropic switched Claude Code's default effort from High to Medium. The reason was that High effort sometimes thought so long the interface looked frozen. Anthropic's internal evaluations found Medium gave slightly lower intelligence with much lower latency on most tasks, and it helped users get more out of their limits. Users then reported that Claude Code felt less intelligent. Anthropic reversed the change on April 7, saying users preferred higher intelligence by default and choosing lower effort themselves for simple tasks. After the reversal, Opus 4.7 defaulted to xhigh in Claude Code and other models to High.
That episode puts the How-To Geek recommendation in context. Lower effort measurably saves usage. Anthropic also found a small but real quality cost, and enough users noticed it in coding work that Anthropic changed course. For the everyday tasks in the How-To Geek test, the cost didn't show up. For agentic coding in Claude Code, Anthropic's own customers decided the default should stay high.
Independent Claude Code Testing Puts Effort Savings Well Below 8x
At least one other published test measured tokens directly instead of reading a usage meter. It points the same way with a smaller gap. According to a daily.dev summary, testing the same five coding tasks on High versus Medium effort showed High generated about 26,000 output tokens versus roughly 14,300 on Medium, a 45% reduction, with Medium completing all tasks and reaching similar conclusions in open-ended cases.
The two tests aren't directly comparable. How-To Geek compared Low against High in claude.ai chat. The daily.dev-reported test compared Medium against High in Claude Code on coding tasks. It makes sense that skipping two effort levels saves more than skipping one. The drop from 8x to roughly 1.8x shows how much the size of the saving depends on the task mix, the product surface, and whether you measure a rounded percentage or actual output tokens.
The two results agree on one thing: for well-defined tasks, the lower setting reached the same outcome. MindStudio's effort guide reaches a similar judgment, saying that for well-constrained, simple tasks, the additional thinking doesn't change the output quality. You get the same result at a higher cost. MindStudio sells an agent-building platform, so its guidance is useful context but not neutral testing.
Claude.ai, Claude Code, and Claude Desktop Draw From the Same Usage Limit
For readers who use Claude across several tools, the most important fact here may be the one least often mentioned. Anthropic's usage-limits article says your usage of all different Claude product surfaces (claude.ai, Claude Code, Claude Desktop) counts towards the same usage limit. Leaving effort high in a chat window uses up allowance that a developer might need later in a Claude Code session in a Windows terminal, and the reverse is also true.
Anthropic lists two main ways to stretch that shared pool. The first matches How-To Geek's advice: choose a lower effort level for routine tasks that don't need Claude's most thorough responses. Higher effort uses more tokens. The second gets less attention: turn off apps you've connected when a conversation doesn't need them, and ask Claude not to search the web when you don't need current information. Anthropic adds that tools and connectors are token-intensive, so managing them helps both maximize your available context window and optimize your usage limits.
Anyone who has connected MCP servers, Microsoft 365 or Google Workspace connectors, or web search to their Claude account can use that second lever alongside effort. The How-To Geek test didn't isolate connectors. Anthropic identifies them as a separate cost, though, so a chat with several active integrations may use more than the effort setting alone would suggest.
A previous faster-than-expected drain was caused by a bug, not by effort settings. The same April postmortem traced reports of Claude Code usage draining unusually fast to a caching bug. After a session had been idle for over an hour, the bug stripped earlier reasoning from every following request, causing repeated cache misses. Anthropic fixed it on April 10 in Claude Code v2.1.101 and reset usage limits for all subscribers on April 23. If your usage drops sharply in a way that seems unrelated to what you're asking, lowering effort won't fix the cause. It's worth checking that Claude Code is up to date.
What this means for you: Choosing a Claude Effort Level by Task
Default to Low or Medium for routine, checkable work, and raise effort only for a specific task that needs it. Anthropic's Help Center gives the same advice, and it fits both the How-To Geek result and the independent Claude Code test. The practical difference from Anthropic's framing is where you start. Anthropic sets High as the default for most models because it balances quality and speed. The How-To Geek author starts at Low and moves up only when needed.
The tasks where lower effort is safest are the ones where you can check the answer yourself: formatting, arithmetic, scheduling, summarizing text you've already read, and small script fixes. Keep effort high for work where you can't easily verify the result, or where Anthropic itself recommends more. That includes long-running coding and agentic jobs in Claude Code, multistep tool use, and hard problems where a quiet error would be costly. Anthropic's Claude Code experience shows developers notice the quality difference in that kind of work.
If you want to measure the saving on your own plan, do it more carefully than a single run. Use the same model and the same prompts. Run each configuration more than once. Where you can, record actual token counts, which is possible through the API or in Claude Code, instead of reading a rounded usage meter. Check correctness as well as cost, since a cheaper wrong answer saves nothing.
- The effort selector appears when you click the model name next to the send button in claude.ai and the Claude apps, on Opus 5.5, Fable 5.1, Opus 5, Sonnet 5, Fable 5, Opus 4.7, Opus 4.6, and Sonnet 4.6.
- Anthropic states that higher effort uses more tokens and reaches usage limits faster, and it recommends Low or Medium for routine tasks.
- How-To Geek's 8x figure comes from one Low-versus-High comparison of three prompts measured on a rounded usage meter. A separate Claude Code test found about a 45% token reduction going from High to Medium.
- Claude.ai, Claude Code, and Claude Desktop share one usage limit, so an effort choice in one affects what's left for the others.
- Turning off unneeded connectors and web search is Anthropic's second documented way to make usage last longer.
- Switching to a more capable model instead of raising effort is the How-To Geek author's personal preference, with no published cost comparison behind it.
The main finding is well supported: Claude's effort setting is the most direct control you have over how fast your allowance runs out, and for everyday tasks the extra thinking often doesn't change the answer. The saving is real, but it varies. It's large when stepping from High down to Low on simple chat prompts, smaller when stepping from High to Medium in coding, and it varies by model. Anthropic's defaults keep changing from one model to the next, with Opus 5.5's API default already down to Medium, so check which level your model starts at after each model release instead of assuming last month's default still applies.