SourceVane · Practical AI

Prompt-cache break-even: how many prefix reuses cover the write cost?

A source-backed calculation compares one cache write plus cache reads with ordinary input processing for a fixed 10,000-token prefix.

Direct answer

The decision in brief

Under the stated Claude Sonnet 5 rates, the first cache read after one write more than covers the higher write cost. The result only applies while the prefix remains eligible and reusable.

Two lines compare ordinary input cost with prompt-cache write and read cost across repeated calls. — SourceVane scenario chart. The same values are available in the HTML table, JSON and CSV.

Answer

One reuse crosses the stated break-even point

Under the stated Claude Sonnet 5 rates, the first cache read after one write more than covers the higher write cost. The result only applies while the prefix remains eligible and reusable.

Reusable prefix10,000 tokens

Cache writes1

Pricing checked2026-09-13

Scenario results

Claude Sonnet 5 prefix-input scenario: $2 standard, $2.50 write, $0.20 read per 1M tokens
Total calls	Cache writes	Cache reads	Without cache	With cache	Savings
1	1	0	$0.020	$0.025	−$0.005
2	1	1	$0.040	$0.027	$0.013
5	1	4	$0.100	$0.033	$0.067
10	1	9	$0.200	$0.043	$0.157
20	1	19	$0.400	$0.063	$0.337

At one call, the cache write costs $0.005 more. At two calls, one eligible read produces a $0.013 saving under these assumptions.

Method and evidence

Without cache: N × P × standard input rate. With cache: one P-token write plus (N − 1) eligible P-token reads. Output and non-repeated input are held equal and omitted because they cancel in this comparison.

Sources: Claude pricing for the stated rates; prompt caching documentation for eligibility and lifetime behavior.

Download JSON · Download CSV · Use the caching decision guide

Limits

This combines source analysis with a deterministic scenario calculation. SourceVane did not run paid Claude requests for this study. Real savings depend on eligible prefix reuse within the applicable cache lifetime, model-specific pricing, cache misses, output tokens and any separate storage charge.

Revision and production notes

Study type: scenario analysis source analysis. Revision: 2026-09-13.1. Updated: 2026-09-13.

SourceVane publishes the assumptions, formulas and machine-readable rows so the result can be checked independently. Corrections change the revision and are recorded on the public page.

Continue this decision

Evidence and next steps

Review the evidence