How much does prompt caching save?

On the rates listed here, a cached read costs between 2.5% and 10% of the normal input price, while writing the cache costs slightly more than normal input. Savings depend on how much of each prompt is a stable prefix and how often it is reused within the cache lifetime.

Are these prices current?

Each rate shows the date it was checked against the provider's own pricing page. Prices change often, so treat the result as a method and confirm the rate before quoting a number in a business case.

DECISION TOOL · FREE TO USE

AI cost calculator

Cost per call, per successful task and per month, with caching, routing and human fallback.

THE WORKING MODELC_task = C_call × calls × (1 + retries) ÷ success

Change an input. See what follows.

Volume

Daily active users

Tasks per user per day

One model call

Cacheable prefix

tokens

System prompt, tool definitions, fixed context.

Fresh input

tokens

User turn, retrieved passages, history.

Output

tokens

Calls per task

Prefix cache hit rate90%

Models and routingMain modelSmall model for easy tasks

Share routed to the small model0%

Only route what your evals show the small model handles. Routing that lowers resolution rate can cost more than it saves.

US-only inference (Anthropic 1.1x)

Outcome

AI resolves the task80%

Extra calls from retries10%

Human cost per escalated task

Illustrative. Use your fully loaded cost.

Retrieval and tools per task

Price per task (for margin)

Cost per call$0.010$0.015 on Claude Sonnet 5.5 with no caching

Per successful task$0.058AI spend ÷ tasks the AI finished

Effective per task$0.846AI attempt + expected human cost

Gross margin44%at $1.50 per task

Per month · 1,800,000 tasks

$83Kinference and tools

$1.44Mhuman fallback

Human fallback is 17.4x the model bill. A one-point gain in resolution saves about $72K a month, more than most model optimisations.

What moves AI cost most · each driver 20% worse

Calls per task+19%
Uncached input+11%
Output tokens+6%
Retry rate+2%
Retrieval and tools+1%

List prices checked against each provider's pricing page on 6 November 2026. Batch discounts, long-context surcharges and cache storage fees are not modelled. Treat totals as a method, not a quote.

FROM THE RESULT TO THE DECISION

How to reason about it

Cost per token is an infrastructure number. The figure a business case is built on is cost per successful task: everything the AI spent, divided by the tasks it actually finished, plus what the failures cost when a person picks them up.

Set the volume, the shape of one model call and the outcome you expect. The calculator prices each call with prompt caching, blends a large and a small model if you route, and shows which input moves the bill most, so you know where an hour of optimisation pays.

Read the human-fallback figure before the model bill. When people handle a meaningful share of tasks, a one-point gain in resolution usually saves more than switching to a cheaper model.

Questions about this tool

Input tokens are processed together in one pass. Output tokens are generated one at a time, each needing its own pass through the model, so they occupy the hardware far longer. Providers price output at several times input to reflect that.