A cache pays from the first re-read, not from day one
Published 18 November 2026 · cache multipliers read from vendor price pages the same day
Caching is not a discount on input. It is a second price schedule: writing the prefix costs more than plain input, reading it costs a fraction. Anthropic charges 1.25× base input for a 5-minute cache and 2× for an hour; OpenAI publishes separate 30-minute and 12-hour write rates. Reads cost 0.1× on most Claude models, 0.05× on Opus 5.5 and Sonnet 5.5, 0.025× on Fable 5.2 — and on OpenAI, roughly 5% of input on the 6.x line.
The question is not whether caching saves money, but how many re-reads the write needs to earn itself back.
Per million cached tokens, Claude Sonnet 5.5
Base input $2.00, cache read $0.10, 5-minute write $2.50, 1-hour write $4.00. Cost of carrying one million tokens of prefix across the first request plus N re-reads:
| Re-reads after the write | No cache | 5-minute cache | 1-hour cache |
|---|---|---|---|
| 1 | $4.00 | $2.60 | $4.10 |
| 2 | $6.00 | $2.70 | $4.20 |
| 5 | $12.00 | $3.00 | $4.50 |
| 10 | $22.00 | $3.50 | $5.00 |
| 100 | $202.00 | $12.50 | $14.00 |
The arithmetic behind the first row: 2.50 + 0.10 = 2.60 against 2.00 × 2 = 4.00. The 5-minute cache wins after one re-read; the 1-hour cache needs two, because its write premium is 1.00× base input rather than 0.25×, while each read saves the same 0.90×.
Break-even as a formula
In units of base input price, a write that is read N times breaks even when N > (write − 1) / (1 − read):
- 5-minute write at 1.25×, read at 0.1× → 0.25 / 0.90 → 0.28, so any re-read.
- 1-hour write at 2×, read at 0.1× → 1.00 / 0.90 → 1.11, so the second re-read.
Where the model changes the answer
The read multiplier is not uniform. On Claude Fable 5.2 a hit costs 0.025× of base input, so each read saves 0.975× and the 5-minute write premium of 0.25× is recovered before the first re-read ends. On models where the hit is 0.1×, the same write takes a full re-read to pay back. The write price is the reason a cache warmed once and never reused is a pure loss: you paid 25% or 100% more for nothing.
Two ways caching quietly fails
- A changed prefix invalidates it. Anthropic's documentation notes that changing the thinking configuration or the top-level effort level between requests can invalidate cache breakpoints; the write is charged again on the next call.
- Batch and cache are separate discounts. No vendor publishes a batch cache-read rate, so combining them is an estimate, not a price. The calculator labels that line.
Model your own cache ratio — set the cache-read share and cache-write volume, and the receipt shows the write and read lines separately.
Questions this note answers
- How many re-reads does a prompt cache need to pay for itself?
- On Anthropic's multipliers a 5-minute write at 1.25× base input pays back after one re-read and a 1-hour write at 2× after two. In units of base input price, a write that is read N times breaks even when N > (write − 1) / (1 − read).
- What does one million cached tokens cost on Claude Sonnet 5.5?
- With base input $2.00, cache read $0.10, 5-minute write $2.50 and 1-hour write $4.00 per million tokens: two requests without a cache cost $4.00, the same pair with a 5-minute cache $2.60 and with a 1-hour cache $4.10. At ten re-reads it is $22.00 without caching against $3.50 and $5.00.
- When does caching lose money?
- When the prefix changes: Anthropic notes that altering the thinking configuration or the top-level effort level between requests can invalidate cache breakpoints, and the write is charged again on the next call. A cache written once and never re-read is a pure loss, and combining batch with a cache read is an estimate because no vendor publishes a batch cache-read rate.
Sources
platform.claude.com/docs/en/about-claude/pricing — cache multipliers, write/read table, break-even statement. developers.openai.com/api/docs/pricing — cached-input and cache-write columns. Read 18 November 2026.