# I Asked Claude the Same Question on Low, High and Max. Only One Looked at My Project. > Source: > Published: 2026-11-06 07:50:43+00:00 Max read my answer back to me. I’d asked Claude a simple question: *“How should I market this app?”* The app is [Savyit](https://play.google.com/store/apps/details?id=com.moneylens.money_lens), a money tracker that reads your bank SMS, so you don’t have to type in your expenses. I asked the same thing three times, on Low, High and Max effort, same model and same project. Low gave me a marketing plan. High gave me a better one. Max didn’t write a plan. It went into my project, found the growth plan, launch post and store brief I’d already written, and basically told me: *your strategy is already here, don’t start over.* Here’s what each one cost: For months I’d either left the effort setting alone or cranked it to Max when I wanted the “best” answer, and I couldn’t feel the difference. That bugged me. Now I think I get it. Low gave me a clean, generic plan: positioning, channels, a 30-day rollout. It answered as if my project were an empty folder. High was sharper. It caught that Google Play restricts the SMS permission, which Low missed completely, and that matters for an app like mine. It was even a cent cheaper than Low. But it still had no idea my project already had marketing docs. Max went and looked, in two rounds. First it found my docs. Then it compared what I’d *planned* with what I’d *built*, and found things I’d missed: None of that came from Max knowing more about marketing. It came from reading my project. I assumed effort meant “thinks longer.” It’s broader than that. Effort decides how much work Claude does: how many files it opens, how much it double-checks, and how far it goes before coming back to you. On low effort it would rather ask you for context than go and dig for it. [Anthropic’s own explainer](https://claude.com/blog/claude-model-and-effort-level-in-claude-code) describes it the same way. That’s exactly what I saw. Low and High answered the sentence I typed. Max looked at the project the sentence was about. The most useful idea I took from that explainer is one question to ask when Claude gets something wrong: did it not *try* hard enough, or did it not *know* enough? Max didn’t *write* more. It *did* more. Every time Claude reads a file or runs a command, the whole conversation gets sent to the model again with the result added on. A plain text answer is one trip. An answer that goes and checks things is many trips. Caching is what keeps this affordable: the part of the conversation the model has already seen is billed at a fraction of the price. Max’s session had a 95% cache hit rate across about 1.5M cached tokens. Without that, it would have cost a lot more than $1.56. One question on one project proves nothing, so I looked for people who’d measured this more carefully. Two tests stood out. Same prompt, five levels. Someone [ran the same game-building prompt](https://huggingface.co/datasets/JOHNISKING/claude-effort-cost) on Opus 5.5 at every level, three times each. Average cost: $0.17 on low, $0.26 on medium, $0.35 on high, $0.82 on xhigh and $5.33 on max, which took about 25 minutes. And every game worked at every level. Max wrote more code, not better games. What happens at the top. [Adrian Cockcroft](https://dev.to/adrianco_54/retort-thinking-level-results-what-opus-actually-does-with-the-extra-time-55f7), formerly of Netflix and AWS, tested Opus 5, the model before 5.5. Below max, it mostly wrote more. At max it switched to *revising*: re-reading its own code, re-running its tests, and only then bringing in subagents. On a routine task that bought nothing, at about 25x the cost. But where something was missing, it helped. Test coverage in Go went from 0.73 to 0.88. That’s my result in a different test. Max isn’t one more notch. It’s a different mode. It’s worth it when there’s something to find, and wasted on a task you already know how to do. If you just want a step up from medium, high is the cheap one. xhigh and max are where the bill jumps. To be fair, the official advice is more conservative than mine: stay on the default effort and fix your context before touching the dial. So treat this as my habit, not a rule. Why the fresh chat? A long chat costs more, because every new message drags the whole history along with it. And the model does worse when you explain a task bit by bit. [One study](https://arxiv.org/abs/2505.06120) found an average 39% drop when instructions came in over several turns instead of all at once. (That was on 2025 models, so take it as a direction, not a law.) A finished plan file is the “all at once” version. Can you switch effort mid-chat? On older models, changing effort threw away the cache, so your next message re-read everything at full price. The [current docs](https://code.claude.com/docs/en/prompt-caching) say that on Opus 5.5, with a subscription or API key, the cache is now kept. Switching *models* still resets it. The fresh-chat routine sidesteps the question either way. This one’s free, and it’s saved me the most. I used to run three or four chats at once, each building something different. My usage would jump from 20% to 40% in a blink. Now I do one thing, finish it, then start the next. The limits feel completely different, and I haven’t hit one since. My usage panel backs this up, roughly. Over the last 24 hours, 91% of my usage came from subagent-heavy sessions, 91% from sessions open for 8+ hours, and 89% from sessions above 150k context. The panel says those numbers are approximate and overlap. But the pattern is hard to miss. So the rule has a second half: keep each chat short, too. Finish the task, save what matters to a file, and start fresh. Since Opus 5.5 came out, my five-hour sessions get a lot more done. But by the time I hit the five-hour limit, 10 to 15% of my weekly limit is already gone. That didn’t use to happen. My guess, and it’s only a guess: bigger sessions under the same weekly cap mean each session takes a bigger bite of the week. I haven’t seen published numbers either way. So watch the weekly one. The one rule I’d take from all this: use high effort when you need discovery and judgment, and low effort when you already know what needs doing. How are you using the effort setting? Have you caught Max finding something you’d missed, or burning money on something simple? I’d like to hear about it. Max wasn’t the smartest answer. It was just the only one that looked at what I’d already built. *The app from this experiment is* *Savyit on Google Play**.* [I Asked Claude the Same Question on Low, High and Max. Only One Looked at My Project.](https://pub.towardsai.net/i-asked-claude-the-same-question-on-low-high-and-max-only-one-looked-at-my-project-9b989c49c21f) was originally published in [Towards AI](https://pub.towardsai.net) on Medium, where people are continuing the conversation by highlighting and responding to this story.