TokenSave Data · Dataset

The coffee-break tax: your next Claude Code message after a pause

Take a break of up to an hour and your next Claude Code message costs the same. Take longer, and it costs about 6 times as much. In our runs Claude Code kept the conversation in Anthropic’s prompt cache for one hour. After pauses of 0 to 55 minutes the next message read 42,667 tokens from the cache and cost $0.042; after 65 minutes the cache had expired, 27,460 tokens had to be written again, and the same message cost $0.253.

Measured 2026-10-20 · Claude Code (Opus 5.5) · pauses of 0, 4, 6, 15, 30, 55, 65 minutes · License CC BY 4.0 · Related: Long chat or new chat?

Key findings

• Nothing changed for an hour. After pauses of 0, 4, 6, 15, 30, 55 minutes, the first model call of the next message read about 42,667 tokens from the cache and wrote only a few hundred new ones. The message cost $0.042 on average whether we waited 0 minutes or 55.
• After 65 minutes the cache was gone. The first call had to write 27,460 tokens to the cache again, at twice the normal input price, and the message cost $0.253, about 6 times as much (+507%).
• The cache lifetime was one hour. Every cache write in these runs was a one-hour write (ephemeral_1h in the usage report), not the five-minute default of the Anthropic API. That is why the 6- and 15-minute pauses, which would have missed a five-minute cache, cost nothing extra.
• The bigger the project, the bigger the tax. Here the context was a small game of about 40,000 tokens. The cost of a cold start grows with everything in the conversation, so after a long session (our long-chat measurement reached 200,000 tokens of context and more) an expired cache costs far more.

The next message after each pause

First model call of the next message: tokens read from and written to the prompt cache. Cost is the API list-price estimate for the whole message (Opus 5.5). Mean of 2 runs. Every run is in the CSV.

What to take from it

• Short breaks are free. Lunch is not: if you come back after more than an hour, the first message pays to rebuild the cache.
• If you know you will be away for longer, it can be cheaper to start a fresh, small session for the next task than to resume a huge one.
• On a subscription this shows up as usage-limit consumption rather than a bill.

Method

• Claude Code 2.1.295 with Opus 5.5 on one Windows laptop, signed in with a subscription, claude -p --effort medium --permission-mode acceptEdits, no MCP servers. Turn 1: “Make a brick-breaker game that runs in the browser as a single file named index.html in the current folder. It needs 2 levels, a score display and 3 lives, and it must be playable with the keyboard and the mouse. Write the file and finish.”. After the pause, turn 2 with --resume: “Make the ball red and slightly faster. Edit index.html and finish.”.
• One session per pause length, all started within a minute of each other on 2026-10-20, so the sessions overlapped in time. Cache reads and writes come from the usage of each model call in the stream-json log; cost from total_cost_usd.

Limitations

• The cache lifetime is set by Claude Code and Anthropic and may differ by login type, version or settings; with an API key and a five-minute cache the break-even pause would be much shorter.
• Because the sessions ran at the same time, part of the shared system prompt may have stayed cached through the other sessions: after 65 minutes about 15,925 tokens were still read from the cache. A lone session may pay a little more.
• Codex runs were started too but stalled on file writes when many ran in parallel, so they are left out.

How to cite

Free to use, quote and republish with attribution under CC BY 4.0. Please link to this page.

Embed the chart

Related

• Long chat or new chat?
• What the Claude Code harness costs
• How to save tokens in Claude Code

Questions or corrections: contact@jonhisking.com