Tip Ten Says Start A Fresh Chat To Save Tokens. Continuing Costs 2.5%.

This card is the opposite of almost everything else this page publishes, and it says so in the caption:

“Most people load Claude with everything they own — global instructions, a folder of old .md files, an about-me doc from three months ago — then wonder why the answers feel generic. They’re not generic. They’re crowded.”

“10 tips in the image. They’re not clever — they’re mostly about deleting things.”

Six of the ten are subtractions. Delete your .md folders. Turn every setting off. Delete your about-me file. Use Skills and Projects only. Don’t stack follow-ups. That is a harder thing to post than a stack diagram, and the diagnosis — crowded, not generic — is the sharpest sentence we have read on a card this month.

Two of the ten pull against each other, though, and they happen to be the last two.

Tip 9 picks the model. Tip 10 throws away its best feature.

Tip 9: “Choose the model. Fable 5.1 on high effort for deep work. Opus 5 for quick jobs.”

Tip 10: “Save your tokens. Don’t stack follow-ups. Start a fresh chat instead.”

Read on its own, tip 10 sounds obviously right: a long conversation resends its whole history every turn, so a fresh chat must be cheaper. That is true of the raw token count and false of the bill, because of caching.

“Cache read tokens are 0.1 times the base input tokens price.”

“Claude Fable 5.1 and Claude Mythos 5.1… priced at 0.025x the base input price.” Claude Opus 5.5 and Claude Sonnet 5.5 are at 0.05x; “all other models use the standard 0.1x multiplier”.

“By default, the cache has a 5-minute lifetime. The cache is refreshed for no additional cost each time the cached content is used.”

“Long multi-turn conversations” is listed among the cases caching is especially useful for.

So on the exact model tip 9 recommends, the conversation you are being told to abandon is the cheapest context in the system to carry.

But the card’s own caption has the right reason

Here is where this stops being a correction and becomes an improvement, because the caption already supplies the better argument for tip 10. “They’re not generic. They’re crowded.”

That is the real case for starting fresh, and it has nothing to do with tokens. A conversation that has drifted three topics away is carrying material that will shape the next answer, and no cache discount fixes a worse answer. The question is not what does it cost to carry this? but is any of it still true?

So the rewrite is a swap of justification, not of advice:

That last row matters more than it looks. The cache is five minutes by default, and it refreshes free every time it is used. A conversation you are actively working in stays warm indefinitely. One you come back to after lunch is cold, and continuing it carries the old context at full price — at which point tip 10 is simply right.

And tip 9 asks you to set a default

Two smaller things in the same tip. “Fable 5.1 on high effort” — high is already Fable 5.1’s default effort. The documentation’s own line for that model is “Start with high, the default.” So there is nothing to change, and that is just as well, because the caching page lists output_config.effort among the things that invalidate the message cache. Changing effort mid-conversation throws away the discount the section above is about.

And “Opus 5 for quick jobs”. Opus 5 is on Anthropic’s legacy list — the current one is Opus 5.5 — and it defaults to high effort too. More to the point, Opus is the heavyweight tier. If the job is quick, the quick models are Sonnet 5.5 and Haiku 5.5, and Haiku starts at $0.50 per million output tokens against Fable 5.1’s $50. We went through the whole ladder in the piece on effort levels.

The eight tips that are just right

It would be a poor reading of this card to spend the whole piece on two tips, because the other eight are the good part and several are genuinely uncommon advice.

One page, two opposite cards

A month after this one, the same page published a stack diagram with twenty-one boxes in it — plugins, skills and MCP servers, four of the boxes advertising 340 items between them. This card opens with “Delete your .md folders” and “Turn every setting OFF”.

Both can be right, and the resolution is not a compromise. They are advice for two different rooms. A stack is what you build when you have a repeated job and know which tools it needs. An empty room is what you want when the job is new and you cannot yet tell which of your old files is making the answer worse. The mistake is running the stack card’s advice in the empty-room situation, which is exactly the failure this caption describes: crowded, not generic.

And the documentation agrees with this card on the mechanism. An enabled plugin “is part of every session, not only the sessions where you use it”, with each skill’s name and description in context “on every turn”. Tip 1 and that sentence are describing the same cost from two directions.

Four honest caveats

1. Cowork is a subscription, not a meter. Those cache multipliers are API prices, and a Cowork user never sees a per-token bill. They are the best public proxy for what consumes a usage allowance, and the ranking holds; the exact ratios are indicative.
2. We did not measure a conversation. This is the card read against the prompt caching, effort and model pages in October 2026.
3. Cache behaviour has conditions. Five minutes by default, refreshed free on use, and invalidated by changes to tools, images, thinking configuration or effort. A conversation you keep poking stays warm; one you leave does not.
4. This is the best card we have taken apart this month. Six of ten tips are deletions, the diagnosis in the caption is correct, and the footer carries a dated check — “checked Sep 2026” — which is rarer and more useful than “features change often”.

The one-line version

Tip 10 is the right instruction with the wrong reason: on Fable 5.1, the model tip 9 picks, carrying a conversation costs 2.5% of sending it fresh and the cache refreshes free every time you use it — so start a new chat when the old one has gone stale, which is what the caption said in the first place.

Keep the eight deletions. Swap tip 10’s reason for the caption’s. Leave tip 9’s effort setting alone, because it is already set.