OpenRouter launches Batch API across 70+ models for cheaper, waitable jobs

OpenRouter has turned on Batch API support for more than 70 models, giving users a cheaper way to run work that does not need an instant answer. The pitch is simple: send a batch, let the provider pick a time within a 24-hour window, and pay about half the usual per-token price — sometimes less.

The company says the slowest possible outcome is not the typical one. During a two-week beta, it saw more than 230,000 completed batches, with a median finish time of 7 minutes and 90% done within an hour. The 99th percentile stretched to 10.3 hours, but most jobs landed far sooner than the full day.

This is aimed at the kind of chores that do not deserve a live request-response loop: labeling text, back-filling embeddings, scoring eval sets, summarizing ticket backlogs, or running one prompt across thousands of rows overnight. The API accepts chat completions, responses, messages, and embeddings, and completed batches return their results inline. Users submit to api/v1/batches and poll GET /api/v1/batches/:id until the job is completed, failed, expired, or cancelled.

OpenRouter also says timing matters more than size in some cases. Batches submitted between 5am and noon Pacific were slower than other hours, while submissions after 6pm Pacific had a 90th percentile under 50 minutes. Even a single-request batch finished in 5 to 11 minutes depending on the hour, and batches of 1,000 or more requests completed in 12 to 21 minutes. For batches over 100 requests submitted between midnight and noon Pacific, the slowest tenth could take as long as 6.8 hours.

There are some guardrails. Each batch runs on a single provider, results come back per request so one bad row does not sink the job, and inputs and outputs are kept for 30 days unless the batch is deleted. Images and files must be public URLs, while audio, video, and OpenRouter’s own web search plugin are not supported in batch.