OpenRouter launches Batch API with 50% discount on most models

OpenRouter rolled out its Batch API yesterday, cutting the price of asynchronous workloads in half in exchange for variable turnaround times of up to 24 hours. The service already supports more than 70 models — chat, completions, messages and embeddings — and targets jobs such as corpus labeling, historical embedding backfill, evaluation runs and overnight support-ticket summarization.

How it works

- POST a request list to `api/v1/batches` specifying the desired endpoint format; receive a batch ID.

- Poll `GET /api/v1/batches/:id` until status reads completed, failed, expired or cancelled.

- Each request in a batch returns an independent result, so a few failures do not sink the whole job.

- Inputs and outputs are retained for 30 days or until manual deletion.

- Every batch appears in the Batches tab in logs with model, provider, status and cost breakdown.

Beta numbers: median 7 minutes, not 24 hours

During a two-week beta, more than 230 thousand batches were processed.

- Median completion: 7 minutes.

- 90th percentile: within 1 hour.

- 99th percentile: 10.3 hours.

The dominant speed factor was submission time, not batch size.

- Batches submitted between 5:00 and 12:00 Pacific were significantly slower; the slowest decile took 2–4.5 hours.

- After 18:00 Pacific the 90th percentile dropped below 50 minutes.

- Single-request batches finished in 5–11 minutes; batches of 1,000+ requests in 12–21 minutes.

What’s supported — and what isn’t

- Routing picks a single provider per batch; by default the cheapest endpoint for the model is chosen after applying the allowed-providers list, data policy and BYOK settings.

- Users who supply their own provider key (BYOK) pay only the BYOK fee.

- Web search is billed at the standard rate, not the discounted batch rate.

- Files and images must be supplied as public URLs.

- Audio and video are not yet supported.