Vertex authentication is enterprise-oriented and may become paid or
licensed in a future release.
Configure
config.yaml:
Service-account authentication
Multiple regions or accounts
Use suffixed env vars to register additional Vertex providers:vertex-us and vertex-eu.
Base URLs and modes
Default bases are derived from project and location:- OpenAI-compatible:
.../projects/{project}/locations/{location}/endpoints/openapi - native Gemini:
.../projects/{project}/locations/{location}/publishers/google
aiplatform.googleapis.com, {location}-aiplatform.googleapis.com,
and aiplatform.{location}.rep.googleapis.com, over HTTPS. For any other base
URL, such as a proxy or a Private Service Connect endpoint, set the auth type to
gcp_service_account and save service-account JSON with the provider. Providers declared in config.yaml or environment
variables are not restricted.
Vertex embeddings use Vertex AI native prediction regardless of
VERTEX_API_MODE. In native mode (the default) the
image endpoints are served through the publisher
endpoints (Imagen via predict, Gemini image models via generateContent);
in openai_compatible mode image generation forwards to the OpenAI-compatible
endpoint and image edits are rejected. Vertex does not expose Gemini Files or
OpenAI-compatible batch operations.
Request translation is shared with the Gemini provider:
the reasoning-effort clamps, the penalties that are never forwarded, and
thought-signature handling apply to Vertex Gemini models
(google/gemini-4-argon, …) in both modes, and native mode also drops the
sampling parameters on Gemini 3 and 4. Gemini 4 is served from the global,
us and eu locations.
Claude models
Setapi_mode: anthropic to serve Claude (Anthropic partner models) from
Vertex. Chat and Responses requests are sent as Anthropic Messages to
.../publishers/anthropic/models/{model}:rawPredict (:streamRawPredict when
streaming), authenticated with the same ADC or service-account credentials.
The base URL may be the /endpoints/openapi or /publishers/google URL you
already use; GoModel derives the /publishers/anthropic URL from it. Vertex
does not list Claude models per project, so list them under models. Use the
model ID from the Model Garden card.
models; keep a second vertex instance for Gemini chat. Token counting and Anthropic batches are not supported.
Image input
Not yet integrated
- Automatic fetching of remote
image_urlvalues in native mode. - Vertex Files and batch prediction APIs.