Prompt Caching + Batch Savings Calculator
See how much prompt caching and the batch API cut your bill on a repeated workload.
If you send the same long system prompt over and over, prompt caching lets the provider bill the repeated part at a much cheaper cached rate. The batch API adds a second lever, taking about half off calls you can run without an instant reply. This prompt caching savings calculator shows how much prompt caching saves and stacks the batch discount on top: enter your tokens and call volume, set how much of the prompt is cacheable, your expected hit rate, and how many calls can go through batch. It reads a shared price table and gives baseline versus optimized cost side by side.
Read the guide: How Prompt Caching and Batching Cut API CostsYour workload
Saved with cache + batch
$948.00
37.9% off $2500.00 on GPT-4o
- Baseline cost
- $2500.00
- Optimized cost
- $1552.00
- Saved by caching
- $560.00
- Saved by batch API
- $388.00
Caching prices the hit share of your cacheable input at the model's cached rate; the batch API applies a typical 50% discount to the batchable share of calls, stacked on top of caching. Real savings depend on your actual cache hit rate and which calls tolerate batch latency. Prices updated January 2026 — verify against the provider before you budget.
How it works
- 1
Set the model and workload
Pick your model, then enter the input and output tokens per call and how many calls the job runs. The model's input, output and cached rates show underneath.
- 2
Tune caching and batch
Slide the cacheable share of your input, the hit rate you expect, and the share of calls you can send through the batch API. Cache is applied first, then the batch discount stacks on the reduced per-call cost.
- 3
Read the savings
The headline is total dollars saved and the percentage off baseline. The stats split the saving into the part from caching and the part from the batch discount, so you can see which lever pays off more.
Instant & 100% private — nothing is uploaded
Every calculation runs locally in your browser. The prompts, token counts and numbers you enter stay on your own device and are never sent to a server — nothing is stored, logged or shared.
Frequently asked questions
- How much does prompt caching save?
- It depends on how much of your prompt repeats and how often it hits cache. Cached input tokens are billed at the model's cached rate, often a fraction of the normal input rate — on GPT-4o it is half, on Claude it can be a tenth. If most of a long system prompt is cacheable with a high hit rate, the input bill drops sharply.
- How does the batch API discount stack with caching?
- The batch API applies roughly a 50% discount to calls you submit for asynchronous processing. This tool applies caching to the per-call cost first, then takes the batch discount off the batchable share of calls, so the two savings compound rather than overlap.
- What is a realistic cache hit rate?
- It varies with your traffic. A shared system prompt reused across many requests within the cache window hits often; a prompt that changes every call rarely does. Caches also expire after minutes of inactivity, so bursty traffic sees fewer hits. Enter the rate you actually observe rather than assuming 100%.
- Why does it say my model has no cache rate?
- Some models in the table have no published cached-input price, so caching cannot save anything for them and the tool flags it. Only the batch discount applies in that case. Pick a model with a cache rate to see the caching side of the math.
Important
For planning and estimates only. Prices come from a published rate table dated on the page; providers change pricing without notice, and token counts here are approximations. Confirm against the provider’s own pricing before you budget or commit.
More tools
More from the Hivly network
Free sister tools on our other sites.
