Batchrouter routes async AI batch jobs to the cheapest eligible provider. We've shipped a lot in the last few months — workflow products, the agent registration flow, OpenAI-compatible endpoints, the credit settlement pipeline — and a fair amount of it never made it into anything readable. This blog is the fix.
Three things will live here, on a rough weekly cadence:
- Engineering notes — how the routing engine actually picks a lane, what the retry policy does on a 429, why we keep one customer batch ID even when execution splits across providers.
- Routing economics — what an OpenAI Batch credit really costs, how usage-based receipts work, when a workflow product beats a direct route.
- Product updates — launches, deprecations, SLA reports, and the occasional incident write-up.
Why we're writing this in the open
A routing platform is a thin layer on top of a lot of opaque choices: which provider, when to retry, when to split a job into lanes, how to settle a sub-cent charge. Most of that complexity gets hidden behind an estimate and a receipt — which is the whole point. But hiding it from operators is wrong. If you ship batch jobs through us, you should be able to read the same notes our team reads.
We'll also be honest about trade-offs. There are things we don't do well yet — fine-grained per-key cost dashboards, on-prem provider lanes, native support for a few research labs' batch APIs — and we'll write about those too, with timelines when we have them.
How to follow along
- Subscribe via the RSS feed or the JSON Feed.
- Share a post with your team — every post has X, LinkedIn, and Hacker News share buttons at the bottom.
- Reply to any post by emailing hello@batchrouter.com — we read everything, and good replies sometimes turn into follow-up posts.
First batch of posts
We're seeding the blog with six posts covering why we built Batchrouter, how we route a job in under 50ms, how to cut your AI bill without changing your prompts, an honest OpenAI-vs-Anthropic batch comparison, lessons from routing 100M tokens, and the case for usage-based receipts in agentic commerce. Pick whichever sounds most useful — they're all standalone.
Onwards.