If you've used the Claude or OpenAI APIs before, you know the obvious advantage they offer: speed. The provider keeps capacity ready for you at all times. That readiness comes at a cost you pay on every call, even for work that could just as easily finish overnight. Your test runs, your data backfills, or your bulk summary jobs — none of them care if the reply comes back in two seconds or two hours. Yet you are still paying for speed you never need.

Both OpenAI and Anthropic will sell you the slower option at half price. Give them a 24-hour window instead of asking for an instant reply, and the same prompt, same model, same tokens cost you 50% less.

You're paying full price for work that could cost half The fast lane has a built-in markup

Some of your API bill is made of live traffic, where a person typed something and is now watching a blank screen. The rest is quiet work that runs on its own. Tests that check whether your prompts still behave, old records being re-processed, thousands of support tickets getting sorted into categories, long documents being summarized for a report that someone will only open next week.

Right now, both kinds of work go through the same door and cost the same money. Fast responses cost more because the provider has to keep machines free and ready for you at all times. When you send a background job through that same fast lane, you are buying that readiness and then throwing it away. It follows the same logic as how you're probably overpaying for Claude without realizing it.

Batch APIs give you a second door. Anthropic calls its version Message Batches, while OpenAI calls it the Batch API. This way, you send your jobs together, the provider runs them whenever it has spare room, and you get the results back within 24 hours. The work itself does not change in any way. Same model, same prompt, same tokens, same quality of answer. The only difference is that you stopped demanding it right now, and for that, both companies cut the price in half.

The money involved is not a rounding error, either. If a third of your spend is background work, you are handing over an extra 15 percent of your entire bill for speed that nobody notices. Many teams are further off than that, because the heavy jobs that chew through millions of tokens are usually the ones running unattended.

Check if your jobs can wait a day (most can) One question sorts your entire workload

Before you move anything, go through your workloads once and ask: is a human waiting on this? If yes, it stays in the fast lane. A chatbot reply, a coding assistant, live document Q&A, anything where someone is watching the screen has to stay synchronous, because any unnecessary delay would break the experience.

Everything else is a candidate. Evaluation runs that score your prompts overnight, backfills that re-process last year's records with a newer model, classification jobs that sort tickets, reviews, or leads into buckets, summaries feeding a weekly report, or embedding and moderation sweeps over content that already exists. None of these have a person on the other end, which means the speed premium you're paying on them buys nothing.

The grey areas usually fall on the batch side too. A report that lands in inboxes at 9 AM does not need to be generated at 8:59; it needs to be submitted the evening before. In practice, the wait is rarely painful; Anthropic says most batches finish within an hour, and the 24-hour figure is the ceiling, not the norm.

Still, plan for the worst case. A batch that isn't done at the 24-hour mark expires, and whatever didn't finish simply doesn't run. So anything with a hard deadline inside a day needs either a retry plan or a place in the fast lane.

Moving a job to the batch lane A resubmission, not a rewrite

There is no way to reroute a running job. A batch is a separate submission path, so you resubmit the work you'd normally send as individual calls. With Anthropic, you assemble your requests into a list. Each entry gets a unique custom_id and a params object holding the exact parameters you'd send anyway, including the model, the messages, the system prompt, and the token limit. You hand that list to the batches endpoint, and you're done. You can even mix unrelated jobs, ticket sorting and document summaries, in a single batch. OpenAI's flow is nearly identical, except you upload your requests as a .jsonl file first, then create the batch with a 24-hour completion window.

From there, it's a waiting game. The batch comes back with an ID and a processing status, which you poll until it reads ended, then you download the results file. That said, the results can arrive in any order, so match them to your inputs by their customid, never by position. Each request also succeeds or fails on its own, and you're only billed for the ones that succeed.

Streaming isn't supported in a batch, and validation happens asynchronously, so a malformed request only reports its error when the whole batch finishes. Dry-run one request through the regular API first to catch mistakes early. The good news is that the discounts stack: prompt caching works inside batches too, and caching can push Claude's per-task costs down to pennies on top of the 50% cut. The limits are generous as well, with Anthropic accepting up to 100,000 requests or 256 MB per batch and keeping results downloadable for 29 days.

Same task, same models, 50% cost-efficient

Batch processing is a great way to reduce your API costs by 50% for the same amount of work, except it doesn't happen immediately. It also has higher rate limits and offers guaranteed completion within 24 hours. If your workload involves running evaluations, classification of large data sets, embedding content repositories, or queuing offline video-render jobs, batch processing is an ideal fit.