If a job does not need an answer right now, the three big API vendors will charge about half. The discount is simple. The conditions are where the work is. Figures here were read from vendor pages on 18 October 2026.
What is published
| Vendor | Discount | Window and limits |
|---|---|---|
| Anthropic | 50 percent on input and output | Most batches finish in under 1 hour, expire if not done in 24 hours, results kept 29 days |
| OpenAI | "50% cost discount compared to synchronous APIs" | Completion window is 24h only. Up to 50,000 requests and a 200 MB input file per batch |
| Listed at 50 percent of standard for most models | Flex is listed at the same price as batch |
Examples from the pages: Claude Sonnet 5.5 batch is $1 input and $5 output per million tokens, against $2 and $10 standard. Opus 5.5 is $2 and $10. Haiku 5.5 up to 100,000 tokens is $0.05 and $0.25. Gemini 3.8 Flash batch is $0.375 input and $1.875 output.
Two details on the Gemini page matter for planning. Standard prices for Gemini 3.8 Flash and 3.6 Flash are listed as rising on 1 January 2027, and Gemini 4 Argon on 1 December 2026, and the batch prices rise with them.
Batch plus caching
Anthropic's page says cache multipliers stack with the Batch API discount. Its batch guide adds a catch: batch requests run concurrently and in any order, so cache hits are best effort. It recommends the 1 hour cache for batches that share context.
Here is an invented job to show the range. Classify 10,000 documents on Sonnet 5.5, each with a shared 20,000 token instruction block, 1,000 tokens of its own and 200 tokens of output.
| Setup | Estimated cost |
|---|---|
| Standard, no caching | $440 |
| Batch only | $220 |
| Batch plus caching, every request hits the cache | $30 |
| Batch plus caching, 90 percent hit | $69 |
| Batch plus caching, 50 percent hit | $225 |
| Batch plus caching, nothing hits | $420 |
Method: batch prices of $1 input and $5 output, a cache read at 0.05 x $1, and a miss billed as a 1 hour cache write at 2 x $1. These are estimates from list prices. The point is that the best case needs a high hit rate, and with a poor one the cache write cost can erase the gain.
What batch costs you
- Time. Do not use it for anything a person is waiting on.
- Tooling. You package requests into files or batch calls and collect results later.
- Failure handling. Anthropic notes batches can expire after 24 hours under load, so plan to resubmit.
Good fits: evaluations, backfills, bulk classification, nightly summaries. Poor fits: interactive coding agents.
Quick check
Why might batch plus caching cost more than you expect?
Batch requests are processed concurrently in any order, so cache hits are best effort. Each miss pays the write price.
Sources
- Claude API pricing, batch processing, fetched 18 October 2026
- Claude batch processing, fetched 18 October 2026
- OpenAI Batch guide, fetched 18 October 2026
- Gemini API pricing, fetched 18 October 2026
