Batch Discount Savings Calculator

How much of your workload can tolerate an asynchronous window, what the discount on it is worth per month, and what one hour of that wait is worth.

The workload

Requests per month

Input tokens per request

Output tokens per request

Share that can tolerate the waitNightly jobs, backfills, evals, enrichment. Anything a user is watching is not eligible.

The offer

Synchronous input rateUS dollars per 1,000,000 tokens

$ / 1M

Synchronous output rateUS dollars per 1,000,000 tokens

$ / 1M

Batch discountThe percentage off both meters for accepting an async window.

Turnaround windowThe worst-case delay you are agreeing to.

hours

Saved per month by batching what can wait

$1,800

That is what a 24-hour turnaround window is worth to you: $75.00 for every hour of delay you are willing to accept, per month.

Cost per request, synchronous: $0.01
Cost per request, batched: $0.0060
Eligible requests per month: 300,000
Ineligible requests per month: 200,000
Eligible work, synchronous: $3,600
Eligible work, batched: $1,800
Ineligible work (unchanged): $2,400
Total bill before: $6,000
Total bill after: $4,200
Saving as a share of the whole bill: 30.0%
Saving per year: $21,600

What this assumes: the discount applies to both meters equally — check that, because a batch tier that discounts input but not output is worth much less on a generation-heavy workload than the headline percentage suggests. The eligible share is billed at the discounted rate and everything else is unchanged, which assumes you can actually route the two classes of work separately. The turnaround window is a ceiling, not a promise: a batch job can complete in minutes or use the whole window, and any downstream deadline has to survive the worst case. Failed items inside a batch usually still cost, and re-submitting them costs again.

A batch tier prices something unusual: your patience. The provider gets to schedule your work into whatever capacity is idle, and hands back a share of what that flexibility is worth. The question is never whether the discount is good — it is whether any of your workload can genuinely wait, and that is a product question rather than a engineering one.

The number this page exists to produce is the last line of the headline: what an hour of latency tolerance is worth per month. It reframes a scheduling decision as a price. If accepting a 24-hour window on your enrichment pipeline is worth a few thousand dollars a month, that is a strong argument for building the queue. If it is worth forty dollars, it is not worth the operational complexity of having two code paths, two failure modes and two sets of monitoring.

The classification of eligible work is where the estimate usually goes wrong, and it goes wrong optimistically. Anything with a person waiting on it is out. Anything on the critical path of a signup, a checkout or a page render is out. What is left is generally the unglamorous half of the system: nightly re-embedding, backfilling a new field across an existing corpus, running an eval suite, generating summaries that are read tomorrow rather than now. That half is often larger than people expect once they go looking, and it is also the half nobody has costed, because it does not show up in any latency dashboard. Run the numbers before you decide it is not worth it.

Batch Discount Savings Calculator

The reasoning behind this

Related tools