You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
provider_name metadata to customize how a provider/model appears in the unified
OpenAI-Compatible catalog and routing — available on every AI entity (providers, audio, image,
video, embedding, moderation, OCR, budget and search-engine models). Resolution order is endpoint_name → provider_name → entity name, so you can rename how a model is addressed
(my-provider/my-model) without renaming the entity.
Azure OpenAI model option for v1 endpoints — in v1 mode the Azure endpoint is
OpenAI-compatible and the model is a real request-body parameter; the new option sets it (and can
be overridden per request), falling back to the deployment id when empty. Ignored in pre-v1 mode
where the deployment id fixes the model.
_default model selector in provider routing — address a provider as provider-name/_default
to explicitly use its configured default model.
Changed
Mistral — reasoning & structured content normalization — typed-chunk content arrays
(text, thinking, image_url, …) from reasoning models (magistral) and multi-part messages are
now normalized in both full responses and streaming deltas: reasoning goes to reasoning_content, images to images, text to content. Unmapped chunk types are preserved
under content_parts. reasoning_content is now also surfaced in the streamed chunk delta.
Refreshed model price database — updated the bundled LiteLLM-based pricing data, broadening and
correcting cost tracking across many providers and models. A warning is now logged when a
model's price cannot be resolved.
Empty model segments are no longer forwarded as an empty model to the provider.
Fixed
Cost tracking with provider fallbacks — streamed responses are now billed against the
provider/model that actually served the request (the last usage entry) instead of the first in
a fallback chain.
Azure OpenAI streaming — a streaming delta with no role now defaults to assistant instead
of failing.