Shipped, dated, and documented.

Eval datasets. Save a named set of test cases and run any eval against it on every prompt version, automatically.

Scheduled evals. Run quality checks on a cron, not just on deploy, to catch model drift between releases.

Trace search is roughly 4x faster on large projects, with new filters for model, status, and latency band.

Budget caps. Set a hard monthly ceiling per project and get paged before you blow past it.

Cost breakdowns now split by model and by route, so you can see exactly what is expensive.

Token counts for streaming responses no longer under-report on early cancellation.

Self-hosted deployment is generally available. Helm and Terraform, bring-your-own KMS, air-gapped mode.

SCIM provisioning and SAML SSO for Enterprise workspaces.

The TypeScript SDK now traces the Vercel AI SDK out of the box.

One-click rollback. Revert to any previous prompt version without a deploy.

Span timestamps are now monotonic under clock skew across distributed workers.

Resolved a rare race that could duplicate a trace when retried at the client.