Claude Token Optimization Guide
- Source
- x.com
- Author
- kaize
- Date
Context
Claude usage depends on the work carried through a conversation, so counting messages alone is misleading. kaize's guide recommends grouping related requests and reusing project documents. Anthropic's current guidance supports those habits: conversation length, attachments, tools and model choice all affect consumption.
Project reuse works through caching: documents uploaded to a project are cached, and while they remain cached, reusing them counts less against limits than new content. That is narrower than the guide's claim that cached content does not eat into usage. It still counts, and caches expire after inactivity, so returning after a long break counts the content in full again. Keeping recurring documents in project knowledge can therefore make repeated use cheaper. On Pro, Max, Team and seat-based Enterprise plans, Settings → Usage shows how much of the session and weekly limits has been used, a more reliable guide than a fixed rule about how many messages a chat should contain.
- Chat cost is quadratic, not linear, because every turn re-reads the whole history: total ≈ S × N(N+1)/2. At ~500 an exchange, 30 messages burn 232K tokens and message 30 costs about 31x what message 1 did.
- One developer's tracking found 98.5% of his tokens went to re-reading history and 1.5% to producing output. The fix is a fresh chat every 15-20 messages, seeded with a summary you asked Claude to write first.
- Batching beats splitting. Three prompts mean three loads; one prompt with three tasks loads context once, and the answers are often better because the model sees the whole job at once.
- Optional features tax every response whether you use them or not — web search, connectors, Explore mode, extended thinking. The default should be off, switched on deliberately when a first attempt falls short.
- Usage runs on a rolling 5-hour window rather than a midnight reset, so spend decays as the day goes on. Burning the whole allowance in one morning session leaves most of the day's capacity unused.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
“Recently I realized that Claude doesn't count the number of messages. it counts tokens.”
“One developer tracked his usage and found that 98.5% of tokens were spent on re-reading the history. Only 1.5% went toward actually outputting the result.”
“Rule: if you didn’t turn this feature on intentionally, turn it off.”
“Choosing the right model is the most important decision you make every day.”
Checking sign-in…
Loading comments…