Analysis·Prompt caching

A cache entry pays for itself on the second read.

Prompt caching costs 25% extra to write and saves 90% on every read. That means the break-even is 1.39 reads — so if a cached prefix is read twice, it has already paid. Across 34 days it saved me $43,327.

RA
Raunak A.
4 August 2026 · 8 min read · No vendor relationships
Precision machinist gauges fanned open on dark felt under raking light.
Illustration generated with AI. Every figure, chart and table in this piece is measured, not generated.

The verdict

A cache write costs 1.25× the normal input rate; a cache read costs 0.1×. So each read saves 0.9× while the write cost 1.25× — break-even lands at 1.39 reads. Read a cached prefix twice and it has already paid for itself. On my workload the ratio was 57 reads per write, which turned a $49,792 bill into $6,465.

Break-even
1.39reads
My read:write ratio
57:1
Saved over 34 days
$43,327
Share of bill saved
87%

Prompt caching is usually explained as a latency feature that also saves money. On an agentic workload that has it backwards. It is a cost feature that also happens to be faster, and the size of the effect is not small enough to treat as an optimisation.

Here is the arithmetic, and then what it looked like against 9.75 billion real cache-read tokens.

The break-even is smaller than you think

Three prices matter, all expressed as multiples of the model's normal input rate:

What each token class costs, relative to input
Token class Multiple Effect
Uncached input1.00×baseline
Cache write 5-minute TTL1.25×+0.25× once
Cache read0.10×−0.90× every time

On the one-hour TTL the write costs 2× instead, moving break-even to 2.22 reads — still under three.

You pay a quarter extra once, then save nine tenths repeatedly. Divide 1.25 by 0.90 and break-even arrives at 1.39 reads. Since you cannot read a fraction of a prefix, the practical rule is:

If a cached prefix will be read twice, cache it. There is no threshold to agonise over.

What it did to a real bill

Over 34 days my workload produced 170,837,424 cache-write tokens and 9,753,346,376 cache-read tokens. A ratio of 57 to 1 — every prefix written was read fifty-seven times on average, forty times past break-even.

Caching effect · 34 days, published list rates
Scenario Cost vs. cached
With caching what it actually was$6,465.11
Without caching every token at full input rate$49,791.807.7×
Saved$43,326.6987%

The cache-write premium across the whole period totalled about $208. It bought $43,535 of avoided read cost.

The write premium — the thing that makes people hesitate — was $208 across 34 days. Rounding error against what it saved.

The failure mode nobody warns you about

Caching is a prefix match. Any byte that changes anywhere in the prefix invalidates everything after it. Which means it can silently stop working, and your bill quietly grows by up to 8× while nothing errors.

The usual culprits are mundane:

Each one turns a 0.1× read back into a 1.0× input charge, invisibly.

Verify it in one line

  • Check usage.cache_read_input_tokens on repeated calls. Non-zero and growing means it is working.
  • Zero across repeated identical requests means something in your prefix is changing. Diff two rendered prompts byte by byte to find it.
  • Remember total prompt size is input + cache-write + cache-read. A small input_tokens is a good sign, not a small prompt.
  • Don't assume caching is on because you set the parameter — a prefix under the minimum cacheable length silently won't cache, with no error

For a workload like mine, an undetected cache regression would have cost roughly $1,400 a day at list rates. It is worth a monitor.

Method & disclosure

From the same dataset as the 34-day cost study — 381 session transcripts, 23,157 billable calls, 18 June to 3 August 2026, de-duplicated by message ID. Cache-read and cache-write token counts are actual, taken from per-message usage records.

Costs are calculated: published per-model list rates applied to measured tokens, cache reads at 0.1× input and writes at 1.25× on the five-minute TTL. "Without caching" prices every cached token at the full input rate. Multipliers are published rates, not measured ones.

No commercial relationship with any vendor named. No affiliate arrangement on anything in this piece.

What AI actually costs, weekly.

Releases, research and analysis — with the money question asked every time. Five mornings a week, about ten minutes each.

Free · No vendor sponsorship of editorial · Unsubscribe anytime

Prompt caching Agent economics Analysis