There’s a discount available from every major AI provider right now. It’s 50 percent, it’s documented, and it asks for one concession: the answer can arrive later.

Almost no small business uses it. Not because it’s difficult, but because the tools default to instant and nobody stops to ask whether they need instant.

The deal

Anthropic’s Message Batches API states the terms plainly: submit a large set of requests at once and costs drop by 50 percent, with most batches finishing inside an hour and a hard limit of 24. Up to 100,000 requests per batch.

OpenAI’s batch pricing is the same shape. Every model on their pricing page has a batch column at exactly half the standard rate. GPT-6-Sol goes from $2.00 to $1.00 on input, $10.00 to $5.00 on output. GPT-5-Nano drops to $0.025 and $0.20.

Google’s Batch API runs at 50% of standard cost too. Their stated turnaround target is 24 hours, with “in majority of cases, it is much quicker” in the docs.

Which of your jobs qualify

The test is one question: does anything bad happen if the answer arrives in 40 minutes instead of 4 seconds?

Work that doesn’t need to be instant:

  • Overnight invoice and receipt processing
  • Summarizing yesterday’s support tickets into a morning list
  • Tagging and categorizing new records
  • Transcribing and summarizing the day’s recorded calls
  • Cleaning or rewriting a backlog of product descriptions
  • Anything you currently do on a Friday afternoon

Work that does:

  • Answering a customer who’s waiting on the line
  • Anything inside a live conversation
  • Anything where a person is sitting there watching a spinner

Most small business AI use falls in the first group. It’s just running in the second group’s mode, because that’s what the tool does by default.

What it looks like in practice

The pattern is boring, which is the point. Through the day, requests get written to a queue instead of being sent. Once or twice a day, the queue goes out as one batch. Results land in a folder or a spreadsheet, and the first person in the office the next morning has finished work rather than waiting work.

The saving is 50 percent on those jobs. Combine it with using a mid-tier model where a mid-tier model is enough, which is most of them, and the running cost of a repetitive process commonly lands somewhere between a fifth and a tenth of what the same process costs when it’s built the obvious way.

The catch, stated honestly

Batch isn’t free to set up. It needs somewhere to hold the queue and someone to check it ran. On a job you run ten times a month that won’t pay back. On a job you run five hundred times a day it pays back in weeks.

Two genuine limits. Batch results can take up to 24 hours under heavy demand, so don’t build anything time-critical on it. And Anthropic notes that large batches can slightly exceed a configured spend limit, so set the limit lower than your actual ceiling rather than at it.

How to check in ten minutes

Ask whichever tool you use most: does it run requests immediately, or can it queue them? If the answer is immediately, look at what it’s doing overnight. Your AI spend isn’t mostly in the answers you read. It’s in the answers nobody was waiting for, sent at full price.