Batch API: half-price inference by bundling requests
By Steven Van ·
Batches process within 24 hours at roughly half the per-token price, with a median completion time of 7 minutes across 70+ models.
OpenRouter has added a Batch API for sending large volumes of requests at once instead of one at a time. Submit a batch through OpenRouter, and the provider completes it within a 24-hour window in exchange for charging roughly half its normal per-token price, and sometimes less. It works today on more than 70 models.
Despite the 24-hour window, OpenRouter says most batches finish much sooner: across 230,000+ batches completed during a two-week beta, the median finished in 7 minutes and 90% finished within an hour. Submission time affects speed, batches sent between 5am and noon Pacific run slower, while those submitted after 6pm Pacific finish in under 50 minutes at the 90th percentile.
Batches support chat completions, responses, messages, and embeddings, and each result returns independently so a few bad rows don't fail the rest of the job. Images and files must be public URLs, and audio, video, and OpenRouter's own web search plugin aren't available in batch (web search calls that do run bill at standard rates). Inputs and results are kept for 30 days or until deleted. It's suited to work like labeling a corpus, back-filling embeddings, or running the same prompt across thousands of rows overnight.
