Polarison

Batch API Discounts Explained: 50% Off the Work That Can Wait

Which models offer a batch discount, what the discounted rate works out to, and which parts of a workload belong in the batch lane.

Updated September 19, 2026

Every major provider sells the same models twice: once for immediate answers, and once at a discount for work that can wait. The discounted lane is called the Batch API, and it is usually 50% off. If any part of your workload does not need a reply in the next second, this is the easiest money you will save.

How batch pricing works

  • You send a file of requests instead of one call at a time, and collect the results later.
  • The provider promises a window, typically up to 24 hours. In practice most batches finish much sooner, but you cannot count on it.
  • Both input and output tokens are discounted, and the discount stacks with nothing else — a batch request is simply billed at the lower rate.

Which models offer a batch discount

ModelProviderDiscountStandard blendedBatch blended
GPT-6 AstraOpenAI50%$20.00$10.00
GPT-5.6 SolOpenAI50%$8.00$4.00
GPT-5.6 TerraOpenAI50%$4.50$2.25
GPT-5.6 LunaOpenAI50%$0.45$0.225
Claude Fable 5.1Anthropic50%$20.00$10.00
Claude Opus 5Anthropic50%$10.00$5.00
Claude Sonnet 5Anthropic50%$4.00$2.00
Claude Haiku 4.5Anthropic50%$2.00$1.00
Gemini 3.8 FlashGoogle50%$1.50$0.75
Gemini 3.5 FlashGoogle50%$3.375$1.688
Gemini 3.5 Flash-LiteGoogle50%$0.85$0.425

Models missing from this table have no published batch rate. That does not always mean the work has to cost full price — some providers discount off-peak hours or slower service tiers instead, and those terms are listed in the notes on each model page.

What a 50% discount is worth

Take the support chatbot workload we use across this site: 100,000 replies a month, ~1,500 input tokens (system prompt + history) and ~400 output tokens each. At list prices that costs $700.00 a month on Claude Sonnet 5. Move the half of it that is not live chat — nightly tagging, summaries, evaluation runs — to the batch lane and that share is billed at half price.

Work that belongs in the batch lane

  • Classifying, tagging or moderating a backlog of content
  • Generating embeddings, summaries or translations for a whole corpus
  • Evaluation and regression runs against a test set
  • Enrichment jobs that feed a nightly report

Work that does not

  • Anything a person is waiting on: chat, autocomplete, an agent step
  • Requests that depend on the previous answer, since each round trip waits for the whole batch
  • Very small jobs, where the extra plumbing costs more engineering time than it saves

Do not forget prompt caching

Batch and caching solve different problems: batch discounts work you can delay, caching discounts context you repeat. Used together on a workload that has both, they compound. The pricing basics guide explains how each is billed.