Batch API: Save 50% on LLM API Costs
Does batch api save money on LLM API costs?
Yes, for the right workload shape: 50% based on Published batch-API discounts across the 6 providers that offer one (providerProfiles.batchDiscountPct) — asynchronous jobs only, not the synchronous API. It does not apply universally — see the applicability and tradeoff sections below before using it.
How it works
Providers price a batch job — a file of requests submitted together and collected once the whole job completes — below the price of calling the same model synchronously one request at a time. The discount is a flat percentage off every model in the lineup, published per provider, not negotiated per volume. It applies to the full request (input and output tokens both), stacking on top of any prompt-caching discount already in effect on the same call.
Where it applies
Only jobs that can tolerate asynchronous turnaround — batch jobs typically return within 24 hours, not synchronously. Live chat, agents, and anything user-facing in real time cannot use it.
Batch discount by provider
| Provider | Batch discount |
|---|---|
| OpenAI | 50% off |
| Anthropic | 50% off |
| 50% off | |
| xAI | — |
| DeepSeek | — |
| Mistral | 50% off |
| Groq | 50% off |
| Cerebras | — |
| Qwen | — |
| Amazon | 50% off |
| Z.ai | — |
The tradeoff
Batch pricing is a different API shape, not a toggle on the one you are already calling: you submit a file of requests and poll for a completed job, which means retry logic, partial-failure handling, and a result-collation step you did not need on the synchronous endpoint. If your workload already runs on a queue (nightly enrichment, bulk classification, offline eval), that cost is close to free. If it does not, building the batch submission and polling path can take longer to pay for itself than the discount saves in the first month, and a workload that is not truly asynchronous (a user waiting on the other end) cannot use it at any discount.
Worked example
For classification at volume (tiny input in, single-label output out) at 5,000,000 calls/month, GPT-5 Nano runs $165/month at list price and $165/month once verbosity is priced in. See the full ranked table for this workload →