Saving Claude tokens: where your weekly limit really goes

In the middle of work, “limit reached” suddenly appears in the terminal, and the reflex is a bigger plan. We measured on one of our own production repos what the budget is actually spent on for a single codebase question. By far the biggest item is the search for the right files.

English edition of an article from our German magazine. Read the German original

Empty by noon: if you burn your token budget on searching instead of thinking, a bigger plan will not help either. Image: Collective Brain (AI-generated)

Be honest: if you use Claude Code seriously in your daily work, in the repo, in the documentation store, in the project folder, you know the moment. In the middle of work: limit reached. And the knee-jerk reaction is always the same: bigger plan, second account, wait for the reset.

The uncomfortable diagnosis: Your limit is probably big enough. It is just being burned for the wrong purpose.

What really happens with a single question

Ask a typical question about a codebase: “How does our change system work, and where is a change processed?”

What the model then does is rarely visible, because it happens in the background. But it can be measured. On one of our production repos (186 files, around 1.4 million words), a single question like this looked as follows:

• Several grep rounds across the whole repo to find candidates
• Five to eight complete files ended up in the context, including a 6,600-word server file of which three functions turned out to be relevant
• Re-reading of passages already read, because earlier excerpts had dropped out of the context window

The five right files alone, the ideal case that never happens in practice because you have to search first, come to almost 13,000 words, over 17,000 tokens. The basis actually usable for the answer: under 2,000 tokens. Everything else is search effort, on every single question.

Extrapolate that to a working day with 30 or 40 codebase questions, and you know where your weekly limit went. For a sense of scale: Anthropic puts the cost across enterprise deployments at around 13 US dollars per developer per active day and 150 to 250 US dollars per developer per month. For 90 percent of users it stays below 30 US dollars per active day. How this plays out in the subscription model, we broke down in the overview of usage credits and billing (in German).

Why “buy more budget” is the wrong answer

In 2026 this is the same thinking trap as “buy more servers” was in 2010: if the process is wasteful, more capacity mainly scales the waste. A limit twice as large that is 90 percent taken up by searching doubles your search costs, not your results.

Anthropic states how much more a bigger plan delivers as a multiple of Pro (as of 2 November 2026): Max for 100 US dollars a month delivers five times as much per five hours and per week, Ultra for 200 US dollars ten times as much per week, without a five-hour limit for verified accounts. Fable is included in Max and Ultra, up to half of the weekly limit. That is a ceiling within the weekly limit, not additional volume. In the Team plan, a premium seat brings five times a standard seat. The share eaten up by searching stays the same with every one of these factors.

The real lever is a different question: What really needs to go into the context per request?

On Team and Enterprise plans, there is a detail that surprises many managers when the bill arrives: the seat allowance resets in a rolling five-hour window and a weekly window, and Claude Code, the Claude chat and Cowork share the same pot. A developer seat draws considerably more from it than a chat seat, because every round carries file contents, tool calls and multi-step reasoning.

The way of thinking: give the model a map instead of a flashlight

The breakthrough in our setup came from upfront work rather than better prompting or a bigger model: condensing the knowledge about the codebase once in advance, meaning structure, symbols, relationships, communities, and giving the model only the relevant section of this map per question, instead of sending it through the whole building with a flashlight.

Measured concretely, same question, same repo:

The conservative factor of 9 assumes that the raw approach hits the right files immediately. It never does. In real sessions, with misses, broad searches and re-reading, we measured up to 72× more answered questions per token budget. A limit that was empty by noon becomes one that lasts the week.

What Anthropic itself recommends

When this article appeared in July, the condensing was still our own detour. Since then the same way of thinking has made it into Anthropic’s own cost documentation, and in surprisingly literal terms. On the code intelligence plugins for typed languages, it says that a single jump to a definition replaces what would otherwise be a grep plus reading several candidate files. And as an example of a useful skill, the documentation names, of all things, a “codebase-overview” that describes the architecture, important directories and naming conventions, so that Claude does not have to work them out by reading several files.

Four more points from the same source can be used in everyday work right away:

• Measuring is now official. On Pro, Max, Ultra, Team and Enterprise plans, the /usage command shows what share of usage falls on individual skills, subagents, plugins and MCP servers, and flags patterns such as long context or cache misses from a share of 10 percent. /context shows what is currently in the window.
• MCP servers cost less context than before, because the tool schemas are loaded only when needed by default. Classic CLI tools such as gh or aws still remain more economical, because they bring no listing at all.
• A hook can pre-filter output before the model sees it. Instead of reading a log file with 10,000 lines in full, the hook greps for ERROR and returns only the hits, which according to the documentation pushes the context down from tens of thousands of tokens to hundreds.
• Agent teams use about seven times as much as a normal session in plan mode, because each teammate keeps its own context window. If you use them, keep the teams small and shut them down actively.

The subagent read 6,100 tokens of files. You got a 420-token result. That’s the context savings.

Anthropic, Claude Code documentation on the context window, accessed on 1 November 2026

Which effort level Claude uses by default

Besides the search, thinking drives usage. The effort level determines how thoroughly the model thinks an answer through, and Anthropic bills thinking tokens as output. According to the Help Center, you reach your limit faster with a higher level. Which level applies if you set nothing depends on the model and the surface (as of 2 November 2026):

Older models with an effort control run on high in Claude Code, Opus 4.7 on xhigh. In the Claude apps, Sonnet 5.5 is also set to medium according to Anthropic. Which level a model uses there by default is marked “Default” in the effort menu next to the send button.

Thinking cannot be switched off for Opus 5.5, Sonnet 5.5, Haiku 5.5 and the Fable models in Claude Code, so your control is /effort. For small changes such as a rename, the documentation names low; for debugging existing code, high. Two details belong with this. If you change the level in the middle of a session, the cache is kept for Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 with a subscription or an API key; with older models it starts over. And a level you saved earlier with /effort, before Claude Code stored the level per model, does not apply to Opus 5.5, Sonnet 5.5 and Haiku 5.5: they start on medium until you choose a level there yourself.

Three levers you can apply today

Even without our complete setup, three habits get you noticeably more out of your budget:

1. Keep a CLAUDE.md, but keep it short. Architecture, conventions, hard rules, a map of the directories: written once, it saves the model the most expensive first rounds of exploration in every session. Since the summer, Anthropic has recommended an upper limit of around 200 lines, because the file lands in the context in full at every session start, even for completely different work. Everything that applies only to individual workflows therefore belongs in a skill that loads only when called. What that looks like in practice is shown in our overview of Claude skills (in German). Our article on Claude Memory shows what belongs in the CLAUDE.md and where it lives (in German).

2. Targeted reads instead of exploring. Say where something is if you know it (“the function is in server/changes.mjs”) instead of letting the model search. Every search round you avoid saves a four-digit number of tokens. For recurring company knowledge, a well-kept project knowledge base (in German) pays off instead.

3. Subagents for broad searches. When the whole repo really has to be searched, delegate the search to a subagent that returns only the result, instead of tipping the dumps of search hits into your main context. The order of magnitude from Anthropic’s own example is in the quote above.

Then there is the cheapest habit of all: /clear between two unrelated tasks. Claude Code sends the complete conversation history with every request, so a session that has been open for hours drags the whole previous conversation along even for a one-line question. And if you pay through the API rather than a subscription, you will find the price side of this calculation in our calculation of Claude API costs (in German).

These are the basics. They might get you to a factor of 2.

The rest of the way is not on the blog

The complete setup, meaning which map we build, what we use to keep it current, how it plugs into Claude Code, and the full benchmark with methodology and raw numbers, does not go to everyone. We send it by email to subscribers of CB Brainwave, our newsletter on AI, marketing and web for mid-sized companies, which is written in German.

Sign up, confirm the link in your inbox (double opt-in, GDPR-compliant, unsubscribe at any time), and the article with the complete setup lands with you:

Arno Hoffrichter

CTO, Collective Brain GmbH · Hamburg

CTO at Collective Brain. Responsible for the technical delivery of web projects and the SEO architecture.

Get in touch