We ship every week, and we write it down.

Submit up to 50,000 requests and get a file back. No latency guarantee, half the per-token fee, and the same eval and trace behaviour as a real-time call. Built for backfills, catalogue jobs, and anything you were doing overnight with a queue.

Enterprise plans can now pin processing, traces, and backups to one region. Previously traces could land in a second region during failover, which was correct for availability and wrong for several of your legal teams.

We moved model scoring off the request path and into a cache refreshed every 15 seconds. Median latency dropped from 310 ms to 240 ms, and Plinth’s own overhead is now 11 ms at p50. No change required on your side.

Found a bad response in the dashboard? Click Promote and it becomes a golden case in the eval suite of your choice, with the expected output prefilled from a model you pick. Building a real eval suite used to take an afternoon. It now takes about a minute per case.

When a provider failed after the stream opened but before the first token, the retry started from the second token. Roughly 1 in 40,000 streamed calls were affected. Fixed, backfilled in the traces, and there is now a test for it.

gem install plinth. Same interface as the Python and TypeScript clients, including streaming and structured output. Go and Ruby are now both at parity with Python.

Attach a deploy identifier as metadata and the dashboard will split spend by it automatically. Useful for the question “did that release make us more expensive,” which turns out to be the most asked question in the product.

Turn it on and we keep nothing once the response is returned. Traces retain metadata only: model, token counts, latency, cost. Available on Pro and above. This one took nine months and rewrote most of the trace pipeline. There is a post explaining why.