On September 22, 2026, OpenRouter, the AI model API aggregation platform, announced on its official blog the launch of the Batch API: bundle a set of requests and submit them together, and the provider schedules execution within a 24-hour window — in exchange, pricing typically drops to 50% of standard per-token rates, and sometimes even lower. More than 70 models are supported on day one.
What the discount buys: certainty
The core trade of the Batch API is straightforward: you give up "results right now," and the provider gains the freedom to schedule off-peak. You call it by POSTing a request list to api/v1/batches with an endpoint shape — chat completions, responses, messages, and embeddings are all supported. After submission, you poll the batch status until it is completed, failed, expired, or cancelled; results come back inline.
The idea mirrors OpenAI's Batch API, but OpenRouter's version has two differences: first, routing defaults to the cheapest batch endpoint available under your provider allowlist, data policy, and BYOK settings, with each batch pinned to a single provider; second, every request's result returns independently, so a few bad rows never sink the whole job. Inputs and results are retained for 30 days and can be deleted at any time with DELETE.
Two weeks of beta data: faster than you'd expect, but timing matters
OpenRouter published data from 230,000+ batches over a two-week beta: the median finished in 7 minutes, and 90% completed within an hour. In other words, you almost never wait anywhere near 24 hours.
But the hour you submit matters: batches submitted between 5am and noon Pacific are significantly slower, with the slowest tenth taking 2 to 4.5 hours; at other hours the 90th percentile drops under 1.1 hours, and after 6pm Pacific it's under 50 minutes. Batch size matters less: batches of 1,000+ requests finish with a median of 12 to 21 minutes.
What fits, and what doesn't
Good fits: corpus labeling, embedding backfills, eval-set scoring, ticket-backlog summarization, running the same prompt across thousands of rows — any workload that can wait overnight.
Bad fits: real-time chat, tasks needing audio or video input (not supported in batch yet), requests depending on OpenRouter's web search plugin (search calls bill at standard rates, and the plugin itself doesn't support batch); image and file inputs must be public URLs.
What it means for API cost structures
The Batch API turns "latency" into a tradable commodity. For indie developers and small teams, workloads like evals and data processing that used to be too expensive to run at scale now have an official, fixed half-price channel. The cost is a workflow change: submit — poll — collect results, instead of the single request-response round trip. Teams that can accept that rewrite will see their inference bills cut in half.
