II.
ModelTransportProtocol overview
Reference · livemodel-transport:vertex-anthropic-messages
Vertex AI — Anthropic Messages overview
Inspect the raw attributes, linked wiki pages, and inbound or outbound graph edges for model-transport:vertex-anthropic-messages.
Attributes
displayName
Vertex AI — Anthropic Messages
vendor
Google Cloud + Anthropic
specUrl
streamingFraming
sse
toolUseSchema
Wire-compatible with Anthropic Messages: tool calls appear as
`tool_use` content blocks (with `id`, `name`, `input`); tool
results are returned as `tool_result` blocks referencing the same
`id`.
thinkingChannel
content-block
cacheControl
explicit
firstSpecVersion
2024-03-19
currentSpecVersion
vertex-2023-10-16
status
standard
requestBodyShape
POST https://{LOCATION}-aiplatform.googleapis.com/v1/projects/{PROJECT_ID}/locations/{LOCATION}/publishers/anthropic/models/{MODEL}:rawPredict
POST https://{LOCATION}-aiplatform.googleapis.com/v1/projects/{PROJECT_ID}/locations/{LOCATION}/publishers/anthropic/models/{MODEL}:streamRawPredict
Headers:
- `Authorization: Bearer <gcloud-oauth-token>`
- `Content-Type: application/json`
Body — Anthropic Messages request shape with these adjustments:
- `anthropic_version` (required, e.g. "vertex-2023-10-16") instead
of the `anthropic-version` header
- no top-level `model` field (it is supplied via the URL path)
- all other fields (`messages`, `system`, `max_tokens`, `tools`,
`tool_choice`, `temperature`, `top_p`, `top_k`, `stop_sequences`,
`stream`, `thinking`, `metadata`) match Anthropic Messages
responseBodyShape
Non-streaming response (HTTP 200 `application/json`): the full
Anthropic Messages response payload —
{
"id": "msg_...",
"type": "message",
"role": "assistant",
"model": "...",
"content": [ ContentBlock, ... ],
"stop_reason": "end_turn"|"max_tokens"|"stop_sequence"|"tool_use",
"stop_sequence": null|string,
"usage": { "input_tokens": int, "output_tokens": int,
"cache_creation_input_tokens"?: int,
"cache_read_input_tokens"?: int }
}
Streaming (`:streamRawPredict`): SSE response framed identically to
Anthropic Messages streaming.
streamingEventTypes
- message_start
- content_block_start
- content_block_delta
- content_block_stop
- message_delta
- message_stop
- ping
- error
toolCallWireFormat
A `tool_use` content block in `message.content`:
{ "type": "tool_use", "id": "toolu_...", "name": "<tool_name>", "input": { ... } }
toolResultWireFormat
A `tool_result` content block in a subsequent user-role message:
{ "type": "tool_result",
"tool_use_id": "toolu_...",
"content": string | ContentBlock[],
"is_error"?: bool }
errorEnvelope
Non-2xx response, `application/json` — Google API error envelope:
{ "error": { "code": int,
"message": string,
"status": "INVALID_ARGUMENT"|"PERMISSION_DENIED"|"RESOURCE_EXHAUSTED"|"NOT_FOUND"|"INTERNAL"|"UNAVAILABLE",
"details": [ ... ] } }
HTTP status mirrors `error.code`.
cacheControlWireFormat
Per-content-block annotation, identical to Anthropic Messages:
{ "type": "text", "text": "...", "cache_control": { "type": "ephemeral", "ttl"?: "5m"|"1h" } }
Up to four cache breakpoints per request. Cache hit accounting is
reported as `usage.cache_creation_input_tokens` and
`usage.cache_read_input_tokens`.
on Vertex (rolled out after the 5m default).
rateLimitSignaling
On HTTP 429 `RESOURCE_EXHAUSTED`: GCP-standard quota error;
`retry-after` header may be returned. Quota state is observable via
GCP quotas / Cloud Monitoring rather than per-response headers.
reasoningWireFormat
Identical to Anthropic Messages — `thinking` and
`redacted_thinking` content blocks on the response message:
{ "type": "thinking", "thinking": "<text>", "signature": "<opaque>" }
{ "type": "redacted_thinking", "data": "<opaque>" }
Both must be echoed back verbatim on multi-turn tool-use loops or
extended-thinking continuations.
authHeaderFormat
GCP OAuth bearer token, sourced from a Google service account or
`gcloud auth application-default print-access-token`:
`Authorization: Bearer <gcloud-oauth-token>`
Project / region selection is encoded in the URL path, not
headers. No `x-api-key` header.
versioningHeader
No `anthropic-version` request header. Protocol version is
declared in-body as `anthropic_version: "vertex-2023-10-16"`.
Optional `anthropic_beta` opt-in features are passed in-body as
`anthropic_beta: ["<feature>", ...]`.
in-body field name (`anthropic_beta` vs `anthropic-beta`) against
current Vertex docs.
Outgoing edges
realizes1
- layer:3-transport·LayerTransport
served_by1
- provider:gcp-vertex·ProviderGoogle Cloud Vertex AI
spoken_by1
- model:claude-opus-4-7@current·ModelVersionClaude 4.7 Opus
Incoming edges
speaks3
- model:claude-fable-5@current·ModelVersionClaude Fable 5
- model:claude-mythos-5@current·ModelVersionClaude Mythos 5
- model:claude-opus-4-8@current·ModelVersionClaude Opus 4.8