Batch API discounts explained: when half price is real, and when it is a trap

Every major provider would rather run your workload at 3 a.m. than at 3 p.m., and they pay you for the flexibility: batch APIs typically cost half the standard rate. On Anthropic, batch processing takes 50% off every Claude model — Sonnet 5.5 drops from $2/$10 to $1/$5 per million tokens. Google’s Batch/Flex tier does the same for Gemini 4 Argon ($1/$5 instead of $2/$10). The discount is real. The question is what it costs you in latency and engineering.

What “batch” actually means

You submit a set of independent requests, the provider runs them when capacity allows — usually within hours, with a completion window of up to 24 hours — and returns the results as a batch. There is no conversation, no streaming, and no guarantee about which minute your job finishes.

That trade is excellent for:

• Document classification, extraction and tagging over large corpora
• Embedding-adjacent preprocessing and evaluation runs
• Nightly report generation, summarization backlogs, data cleaning
• A/B evaluations of prompts or models over thousands of samples

It is wrong for anything a user is waiting on: chat, agents in a live loop, autocomplete, support replies. A 50% discount on a request that arrives a day late is not a saving; it is an outage.

What the saving looks like at volume

Take a document pipeline on Claude Sonnet 5.5: 50,000 input and 2,000 output tokens per document, 1,000 documents a month.

• Standard tier: (50,000 × $2 + 2,000 × $10) ÷ 1M = $0.12 per document → $120/month
• Batch tier: exactly half → $60/month

At a million documents the same decision is worth $60,000. Batch eligibility is the single largest line item in most LLM budgets after model choice itself — larger than prompt optimization, because it applies to every token unconditionally.

The fine print

• Discounts stack with nothing. Batch rates replace standard rates; cached-input discounts apply on top only where the provider says so.
• Failure handling is yours. A malformed request fails inside the batch, not at submission. Validate inputs before you send a million of them.
• Windows differ. Completion SLAs and maximum batch sizes vary by provider and change quietly; check the current limits before architecting around them.
• Not every provider publishes a batch rate for every model. Where our dataset shows “Not published”, the model page says so rather than assuming the usual 50%.

A simple decision rule

If the result is needed in under a minute, pay standard. If it is needed within a day and the requests are independent, batch — and put the saving in writing, because at volume it routinely exceeds the entire prompt-engineering budget. Every model page on this site shows its batch rate next to the standard one, and the calculator can price either tier.