Every user-visible change to the API, the router, and the dashboard. Breaking changes are announced here first and carry a 12-month deprecation window.
Tag any request with a customer ID and spend, tokens, and latency roll up per customer in the dashboard and in the usage export.
Long-running tool calls could exceed intermediary idle timeouts on some networks. The gateway now emits a comment frame every 15 seconds during tool execution.
Re-run any traced request against a different model or routing policy and diff the outputs, without sending traffic to production.
The scorer now runs on the same host as the gateway, removing a network hop from the hot path.
Shared prefixes are cached at the gateway across requests in a project, cutting input token cost on workloads with long system prompts.
ap-south and sa-east joined the pool, and every region now runs active-active rather than active-standby.
A rewritten scorer that estimates request complexity before selecting a model, rather than relying on static per-endpoint rules.
Automatic user and group provisioning for Enterprise workspaces, alongside the existing SAML support.
JSON-schema-constrained output now works on every open-weight model in the catalog, not only frontier models.
Image tokens were counted at the provider's pre-tiling rate for two vision models, understating usage by up to 4%. Affected accounts were credited.
Older releases are in the archive.