Living Well with Claude Code Limits: Learning 'How Tokens Are Used' from Official Anthropic Tips
Hello, this is Horiuchi from Renewer Inc.
When you use Claude Code extensively, you often find yourself worrying about weekly or 5-hour limits, even on the top-tier Max plan. I myself often hit these limits and have my work stopped, or I frequently have to adjust my usage to avoid exceeding them.
"I want to use Claude Code as much as possible without worrying about limits."
To address this desire, Anthropic has published an official explanatory article. It is a blog post published on August 14, 2026, titled "
Maximizing the value of your Claude Code sessions".
It introduces tips for avoiding hitting limits when using Claude Code and for maximizing the value and productivity of your Claude Code usage. The day after its release, it became a hot topic in overseas practitioner communities like Hacker News.
In this article, I will break down the key points of this collection of official tips. The original text is in English and somewhat technical, so I have re-translated it into a form that is easy for non-engineers to understand.
│
In the "Claude 5 Generation Master Guide" I published the other day, I explained the "6 New Rules" of the Claude 5 generation (what to pass to the AI) and "Model Selection" (which model to pass it to).
It is a document of over 50 pages that covers how to master the Claude 5 generation, including "New Methods of Context Engineering" and a "Claude 5 Model Selection Guide," centered on official Anthropic articles. You can download it for free from the link below.
This collection of tips is a continuation of that guidebook. If the guidebook is the design for "how to pass information," this is the design for "session management." If you read these as a set, you can see the overall picture of how to use it that Anthropic envisions. I will also introduce those connections in the relevant sections.
1. Why is Anthropic starting to talk about the "value of tokens" now?
First, let's start with the premise. Claude Code's paid plans have usage limits. When these limits were introduced in 2025, there were voices of dissatisfaction, especially from heavy users.
This article emerged within that context. I will quote the opening message.
"Being efficient with tokens doesn't mean using fewer of them overall. It means making sure the ones you do use go towards the thing you actually asked for."
(Token efficiency does not mean reducing the total amount used. It means ensuring that the tokens you do use are directed toward what you actually requested.)
They are saying that this is a story about "designing value," not "saving techniques." Even for the same task, if the session state is poor, the AI will ponder over 10 irrelevant files. The tokens spent on that contribute not even a millimeter to your request.
In other words, if you cut out wasted tokens, your actual productivity will increase while keeping the same price and the same limits.
In my previous article, I introduced Anthropic's argument to "shift your accounting unit from 'tokens' to 'tasks'," and you can see their stance of focusing on helping users increase their productivity by focusing more on their work and the value they provide.
2. The 3 factors that determine token price
The first half of the official article explains 'what determines the price of 1 token.' There are three factors.
・Model selection (larger models are more expensive)
・Input and output (output tokens cost about 5 times more than input)
・Prompt caching (reading is 0.1 times the normal rate, writing is up to 2 times)
The third factor, prompt caching, is a mechanism that reuses the premise of a conversation once it has been read as 'saved.' Reading with an active cache costs only 0.1 times, or one-tenth of the cost of a normal input.
This cache seems to break quite easily.
Executing /model (changing the model) or /effort (changing the depth of thought) in the middle of a session causes the entire conversation to be re-read.
It's like having to pay the full price again for something that was costing one-tenth.
Therefore, this is the basic behavior recommended by the official team.
Decide on the model and effort level at the start of the session and do not change them midway.
It's subtle, but once you understand the mechanism, you'll likely want to practice this (and won't want to change the model or effort).
So, how do you decide on a model at the start?
You can use the 'Start Smart' concept introduced in the previous article, 'Choosing a Model for the Claude 5 Generation,' exactly as it is.
First, verify the quality limit with the highest-performance model × standard settings.
If you want to lower costs, lower the effort level before lowering the model. That was the order.
Combined with these tips, it becomes a single operational rule: 'Decide at the start with Start Smart, and do not change it during the session.'
3. Cache has an 'expiration date'
Another point introduced about caching is the expiration date.
The cache expires in 1 hour (for subscription users; it is 5 minutes for API key users). In other words, the first move you make after returning from a lunch break is usually a full-price payment.
Therefore, the official recommendation is to type
/compactbefore leaving your desk.
/compact means 'compression' in Japanese.
It has the AI summarize the conversation up to that point, replacing the long history of exchanges with a short summary. Since the context of what you were doing is carried over, you can continue working as is.
'I'm still in the middle of work, but the conversation has become long'—this is a command to use when you want to keep costs down.
Furthermore, if you execute compression while the cache is still active, you can execute it while keeping the token amount low.
4. Diagnosing 'what to put' in the context
Next is the topic of the content (context) loaded into the session.
Actually, before you even type the first character, tool definitions, CLAUDE.md, and MCP definitions are already loaded.
How much of what is your session carrying? There is a command to visualize that.
/contextWhen executed, a breakdown of the initial load is displayed in a graph.
In the desktop app, clicking the circular 'usage ring' next to the model selection at the bottom right of the input field displays the same breakdown. The following screen is what I opened in my environment.
By the way, you can also jump to the overall consumption status of your subscription plan from the "Plan Usage Limits" section at the bottom. Just by looking at this screen once, you can see what is "heavy" in your environment.
An important action to reduce context is to reduce the number of MCP connections. There are many reports in the community that "initial loading time was halved after organizing MCPs."
Once you can visualize it, perform the reduction work.
The solution is simple: just disable the servers you don't use via the /mcp command.
In the desktop app, MCPs are handled under the name "Connectors."
You can add and manage them from the "+" button next to the input field → Connectors, and you can disconnect ones you want to remove via "Manage connectors" or Settings → Connectors.
5. Deciding "when to discard" context
An important fact introduced by the official team is that
"Even for the same amount of work, doing it in one long session is more expensive than splitting it into shorter ones."The reason is simple: in the 40th turn of an interaction, the AI is re-reading the previous 39 turns. As the conversation grows, the amount of data re-read behind the scenes for each response accumulates.
So, how can you organize that massive context?
When a task is finished, use /clear to erase the context. That's all.
The /clear command discards the entire conversation history of the currently open session. From the AI's perspective, it becomes a state where it cannot remember what you were talking about just a moment ago.
If the trial and error from the previous task remains, the AI will continue to read it as "information to consider" in the next task. This is a phenomenon known as context rot, where accuracy also drops.
I will also organize how to use it properly alongside similar commands.
・/clear ── Erase everything. When switching tasks.
・/compact ── Summarize and compress. When you want to continue working but it has become too long.
・/rewind ── Rewind only the most recent interactions. When the conversation has gone off-track.
A new command called /rewind has appeared.
The official team says, "When you make a mistake, prioritize /rewind over /compact."
"If the last few turns went somewhere you don't want to keep, /rewind to just before them instead of running /compact. Rewinding only cuts those turns off the end, so everything before them is still cached and it costs nothing."
(If the most recent interaction has gone in an unwanted direction, do not run /compact, but instead use /rewind to go back to just before that point. Rewinding only cuts off the end of the interaction, so everything before that remains cached, and the cost is zero.)
In the desktop app, the equivalent of /clear is a "New Session."
The desktop app is designed with the premise of "splitting sessions by task," and each session has its own independent context. In other words, "/clear when a task is finished" is unnecessary in the app; please interpret it as "create a session for each task."
/compact and /rewind can also be used in the desktop app. They work just by typing them into the input field (automatic summarization also occurs when the context becomes full).
6. How did the overseas community receive this?
On the Hacker News thread, opinions were clearly divided.
Supporters appreciated the fact that "the official team disclosed the internal workings of the tool" itself. A trend of bringing in additional tips also emerged, becoming a place for exchanging practical knowledge.
On the other hand, a representative opinion from the critics was this: "Isn't this shifting the burden of cost management onto the user?" This is the stance that the tool side should optimize automatically.
Perhaps automatic optimization by the tool side will also be realized in half a year.
7. Summary ── Priorities to check
The official article concludes by listing four priorities for where to look first, ordered by their impact on cost.
・Session length and context size
・Model and effort level selection
・Whether prompt caching is broken
・Number of turns and tool call frequency
The most important takeaway is to first question "Is this session getting too long?" Personally, this was the part that gave me the most insight.
We tend to use one long, continuous session (because the context is maintained and it's faster to talk), but we need to be mindful of it in terms of cost.
Thank you for reading this far!
Moving forward, I believe that for both personal and corporate use, efficient AI model usage will be essential for achieving ROI. After "writing good prompts," the next skill might be "knowing how to wrap up sessions effectively."
The six new rules of context design and the model selection guide, which form the basis of this topic, are summarized in the "Claude 5 Generation Master Guide." You can download it for free from the link below.
Information on using Claude for work and business is summarized in the magazine below.
Please read it as well. If you like it, please follow me.