How to · Anthropic

Claude Batch API Pricing: Half Price for Work Nobody Is Waiting On

By Ghiles Asmani, founder of Weckr · Published August 15, 2026

Short version:Anthropic’s Message Batches API processes requests asynchronously, typically within 24 hours, at 50 percent off both input and output tokens. Sonnet 4.6 drops from $3 and $15 per million tokens to $1.50 and $7.50. If a workload lands in a database instead of a chat window and you run it at the synchronous price, you are paying double for latency nobody notices.

The batch discount is one of the last large, boring, guaranteed savings in LLM operations: no quality tradeoff, no model change, just a different endpoint for work that was already asynchronous in spirit.

The numbers

Batch pricing is the standard rate divided by two, on both sides. At current published rates (per million tokens, verify on Anthropic’s pricing page):

Model              Standard          Batch (50% off)
claude-opus-4-8    $5.00 / $25.00    $2.50 / $12.50
claude-sonnet-4-6  $3.00 / $15.00    $1.50 / $7.50
claude-haiku-4-5   $1.00 / $5.00     $0.50 / $2.50

Worked example: a nightly job summarizing 10,000 documents on Sonnet, each about 2,000 tokens in and 400 out, costs $120 synchronous ($0.012 per call) and $60 batched. Every night. OpenAI’s Batch API offers the same 50 percent structure, so this section of your architecture is provider neutral.

What belongs in a batch

  • Backfills and migrations: classifying or summarizing your existing corpus.
  • Nightly and scheduled jobs: digests, reports, enrichment queues.
  • Evals and prompt experiments: the version testing from LLM cost per prompt version at half price.
  • Anything queued:if you already show the user “we will email you when ready”, the latency budget for batching exists.

What does not belong: anything interactive, and anything with a hard deadline under the completion window. The 24 hour figure is a typical bound, not a guarantee of speed.

The tracking wrinkle: attribute at collection, price at batch rate

Batched calls still belong to users and features: the nightly digest job is spend caused by specific customers on a specific feature, and it belongs in their margin. Two rules keep the books honest. Carry the user id and feature label on each batch item so attribution survives the round trip. And record batched tokens at the batch rate, because pricing them at the synchronous rate overstates spend 2x, which quietly distorts margins and can trip per user spending caps early, the same class of wrongness as the stale tables in LLM price history, just in the other direction.

Weckr’s log endpoint takes batched events like any other: user, feature, model, tokens, and the cost is recomputed server side. The result is one margin view where a customer’s interactive chat spend and their share of the nightly batch both show up against what they pay.

FAQ

How much cheaper is the Claude Batch API?

Fifty percent on both input and output tokens, in exchange for asynchronous processing with results typically returned within 24 hours. At current rates that puts Opus 4.8 at about $2.50 input and $12.50 output per million tokens, Sonnet 4.6 at $1.50 and $7.50, and Haiku 4.5 at $0.50 and $2.50. Confirm against Anthropic’s pricing page before relying on any figure.

What workloads fit the Batch API?

Anything a user is not waiting on: nightly summarization runs, embedding or classification backfills, bulk document processing, evals, and content generation queues. The rule of thumb: if the result lands in a database rather than a chat window, it is probably batchable, and running it at full price is a 2x overspend.

Does the Batch API work with prompt caching?

The discounts stack conceptually: batch halves the base rates and caching discounts repeated context on top of the applicable rates. For large runs sharing a big system prompt, the combination is dramatically cheaper than naive sequential calls. Check current documentation for the exact interaction on your models, since cache mechanics differ by cache lifetime.

Does OpenAI have the same batch discount?

Yes, OpenAI’s Batch API also gives 50 percent off with a 24 hour completion window, so the batching decision is provider neutral: the moment a workload becomes asynchronous, it should also become half price, whichever provider serves it.

How do I track per user cost on batched calls?

The same way as synchronous calls, with one difference: attribution happens when results return, not when you submit. Keep your user id and feature label on each batch item, and when you collect results, log tokens per item against the batch rate. If you report batched tokens at full price your cost tracking overstates spend 2x, the mirror image of the stale price table problem.

Keep reading

Anthropic Billing Explained

How the Claude API charges you, and what the console hides.

Prompt Caching Economics

The 90 percent discount with a 25 percent trap.

Claude API Cost Optimization

Six ranked levers to cut your Claude API bill.

Halve the queue, then watch the whole picture

Batching is a one time architectural win. Knowing which customers your batches serve, and whether they pay for themselves, is the ongoing one. Weckr tracks both interactive and batched spend per user against plan price, from two lines of integration. See it on the live demo, or start with the AI cost and margin guide.

See the dashboard with real data, no signup needed.

Try the demo →