Interactive tool
Claude Code Cache Cost Calculator
Compare the estimated cost of a warm cache read, a cold prompt-cache rewrite, and a keepwarm window for one return message.
This independent Cache Tax calculator uses API-equivalent price estimates. It runs in your browser and never connects to Claude Code or reads your transcript.
Set up your scenario
Use the cached prefix size from your session if available. All inputs stay in this browser; no transcript is uploaded.
Jump to the estimateAdjust return and keepwarm assumptions
Preset API list prices checked 2026-10-11. Verify current model and provider rates before relying on a result. Official price table
Warm cache return
$0.025
Cached prefix read + new input + output
Cold cache rewrite
$0.805
Prefix write + new input + output
Your break: 90 minutes
Without keepwarm (cold return)$0.805
Estimated pings in window1
Cost per ping assumption$0.0201
With keepwarm, including pings$0.0451
Estimated difference$0.7599 lower
How this estimate works
Each amount is tokens × USD per million tokens ÷ 1,000,000. The model assumes the entire prefix is either read or rewritten once on return. A keepwarm ping reads that prefix and adds the assumed new input and output tokens.
Real requests can have mixed read and write tokens, prefix changes, provider pricing, extra turns and different ping output. Subscription users should read these dollars as API-equivalent values, not an invoice or a prediction of plan limits.
What the three numbers mean
Warm return: the assumed cached prefix is read. Cold return: that prefix is written again at the selected cache-write rate. Keepwarm: scheduled pings add cache reads and output, then the return is estimated as warm if the selected window covers it.
This is a single-request comparison. Actual Claude Code turns can contain several model calls, mixed cache reads and writes, and different context sizes. A change to the prompt prefix can invalidate part of the cache even before the timer expires.
A worked example
At the listed Claude Fable 5.1 rates checked on 2026-10-11, a 200,000-token prefix costs about $4.00 to write to a one-hour cache and $0.05 to read while warm. A ping that reads the prefix also has new input and output costs, so its real total is above the read-only amount.
For subscriptions, these dollar figures are a comparison yardstick, not a charge forecast. Rate-limit effects are not calculated.
Where the rates come from
Preset rates come from the Anthropic prompt-caching price table, checked 2026-10-11. The five-minute write is 1.25× base input; the one-hour write is 2×. Cache-read multipliers have model-specific exceptions, so the calculator stores each model’s listed read rate instead of assuming one percentage for all models.
If your model, provider or contract uses another rate, select Custom rates and enter all five prices. Check why the prefix may need rewriting before treating a cold estimate as certain.