OpenRouter Adds Batch API With Half-Price AI Requests for Non-Urgent Work

OpenRouter has launched a Batch API that turns patience into a pricing lever. Developers can submit large or non-urgent AI jobs for completion within a 24-hour window and generally pay 50% or less of normal per-token rates, across more than 70 models.

The offer is not meant for an app waiting on a user. It is for work that can be moved out of the immediate request path: labeling a corpus, generating embeddings for older material, scoring an evaluation set, summarizing a ticket backlog, or applying one prompt across thousands of rows overnight. The trade is explicit: lower token prices in exchange for a provider-controlled completion window.

A lower bill, with timing no longer guaranteed

That distinction changes the product from a simple cheaper endpoint into a scheduling choice. A normal request is valuable when an answer must arrive now; a batch is valuable when a team can separate the cost of processing from the moment someone needs the result. OpenRouter says the discount varies by model, while web-search calls continue to bill at standard rates.

Fast in beta is not the same as immediate

The early results make the 24-hour ceiling look conservative for many jobs, but they also show why this is not a latency promise. OpenRouter’s two-week beta data puts the 99th-percentile completion time at 10.3 hours. Submission time mattered more than request count, according to the company: batches sent between 5 a.m. and noon Pacific were significantly slower, and the slowest tenth took two to 4.5 hours.

Batch size still has an effect at the margins. OpenRouter says a single-request batch finished in five to 11 minutes depending on the hour, while batches with 1,000 or more requests finished in 12 to 21 minutes. Its slowest tenth of batches larger than 100 requests, submitted from midnight to noon Pacific, took as long as 6.8 hours.

Broad request support, narrower media support

The API is built to fit several existing OpenRouter request shapes rather than impose a new task format. It supports chat completions, responses, messages and embeddings, and results return independently, so failed rows do not cause an entire batch to fail.

Operating limits to account for

• Each batch runs on one provider. By default, OpenRouter selects the cheapest eligible batch endpoint after applying provider allowlists, data-policy settings and bring-your-own-key settings.
• Inputs and results remain available for 30 days unless the customer deletes the batch.
• Images and files must use public URLs. Audio, video and OpenRouter’s web-search plugin are unavailable in batch jobs.

Those boundaries narrow the best fit. The Batch API can reduce costs for text-heavy work with a known queue and a flexible deadline. It is less suited to tasks needing private file uploads, media processing, web-search calls at batch pricing, or a predictable response time. For teams that can wait, the relevant comparison is not between models alone, but between paying for immediacy and paying for volume.