Skip to main content

🌉 Beak AI Gateway – Release Notes

Release notes for model routing, load balancing, and supported LLM integrations.

Version 6.0.0​

A new tenant arrives with the model catalogue in place​

Creating a tenant now sets the Gateway up for it, including the catalogue the model picker lists — every supported model, each provider's credential form, and the alert channel types — plus a default project and team for usage to be attributed to.

  • The model picker is no longer empty for a new tenant. The catalogue used to arrive only by copying an existing tenant's database; a tenant created any other way opened the Gateway to "models data not available" and could not add a model at all.
  • No credentials are ever copied. What is installed is the FORM for each provider — its field names and types — never a value. Adding a model with a real credential stays the administrator's step.
  • A default team and project are created, so spend and rate limits have somewhere to be recorded from the first request.

On by default for newly created tenants; existing tenants keep the catalogue they already have — nothing is overwritten. See What Each Product Sets Up.

Point an OpenAI Agents SDK or Claude Code agent at Beak without changing its dialect​

The Gateway now serves the OpenAI Responses API at /v1/responses and the Anthropic Messages API at /v1/messages, with /v1/models for discovery and /v1/chat/completions as an alias of the native route. An agent built on the OpenAI Agents SDK keeps its default API; Claude Code and the Claude Agent SDK connect with ANTHROPIC_BASE_URL alone. See AI Gateway Integration — Other dialects.

  • Same governance, whichever dialect. Key, budget, model access, rate limits, prompt moderation, routing, spend attribution and the completion audit run identically; only the request and response framing differ.
  • Errors in the caller's own shape. A moderation block is a 403 in the OpenAI or Anthropic error envelope, with the same reason the native route gives.
  • Streaming in both dialects, with spend settled from the stream's final usage as on the native route.
  • Unknown request fields are dropped, not forwarded. A parameter the translation cannot express in Chat Completions never reaches a provider.

Note: Nothing changes for existing /chat/completions callers, and no configuration is needed to enable the new paths — they authenticate with the same virtual key. Tenants with security disabled are not scanned on any surface.

Version 1.0.0​

  • Added multi-provider routing.
  • Model config per project support.