TokenSkim

Batch API vs prompt caching: which saves more?

These two levers are often confused. They target different waste and, crucially, they combine: a batched request with a cached prefix gets both discounts.

Caching targets repetition

Caching saves on repeated input — the same prefix sent many times. If your prompts are large and stable, caching is your biggest lever.

Batch targets latency tolerance

Batch saves 50% on anything that doesn't need an immediate answer, on both input and output. If a lot of your spend is background jobs, batch is your biggest lever.

They stack

You don't have to choose. A non-urgent, repetitive workload should use both — the batch discount applies on top of the cache discount.

Turn this into your number

Drop your usage export into the free analyzer and see how much of this applies to your account — provable savings separated from estimates. Nothing is uploaded.

Analyze my usage — free