90% Cheaper AI Cache Reads Do Not Mean a 90% Smaller Bill

We use Sonnet 4.6’s documented multipliers to keep this calculation specific. Not every Claude model has the same read multiplier. The current pricing table lists exceptions, so use the row for the model you actually run.²

You are reusing the briefing, not buying the answer twice

Picture a support assistant that receives the same long company policy with each customer question. Prompt caching lets an eligible repeated beginning of the prompt be reused instead of charged as ordinary input each time. It does not mean recycling the previous customer’s answer.¹

That distinction explains the opportunity. A large, stable policy used repeatedly may be a useful candidate. A short instruction that changes on every request may not be. The question is how much input really repeats, and how soon.

The first repeat can make the five-minute option pay

Let one unit mean the price of sending the reusable portion once without caching. The table shows only that portion’s input cost. Request counts are hypothetical, using the documented multipliers above. Every later request is assumed to get a cache hit.

With one later hit, the five-minute option costs 1.25 + 0.10 = 1.35 units instead of 2. That is a 32.5% saving on the repeated input across the two requests. The one-hour option still costs more at that point. It becomes cheaper than no caching after its second hit.

The general break-even check is simple: write cost + number of hits × read cost must be below the uncached cost of those same requests. Count a fresh write whenever the relevant cached input has to be stored again.

Timing can erase the saving

Anthropic documents a five-minute default lifetime that refreshes when the cache is used, plus an optional one-hour lifetime. The cache also has matching and minimum-length requirements. A repeated prompt does not automatically guarantee a hit.¹

A burst of similar requests is different from one request every few hours. The burst may reuse a single write. The spaced-out requests may repeatedly pay to create a cache that nobody uses again. Average daily volume alone hides that difference.

Other providers need their own calculation. Gemini’s caching documentation distinguishes implicit and explicit caching, and explicit caching has storage-related pricing as well as token costs. Do not transplant Claude’s break-even table into a Gemini budget.³

Measure the share of the bill that can actually fall

Suppose reused input accounts for half your model bill and a change reduces that part by 50%. The total bill falls 25%, not 50%, if everything else stays the same. That is illustrative arithmetic, but it is the right way to think about a discount on one part of a request.

Use actual usage records to compare cache writes, reads and uncached input. Then place the saving inside the cost of finishing the task. A smaller prompt bill is useful. It is not the whole business case.