Verified September 2, 2026
Claude Fable 5.1 on Vertex AI
If your data already lives in GCP, Vertex is the natural front door — but Fable 5.1's quota, region list, and 30-day retention rules are not the same as calling Anthropic directly.
The short answer
Available in Vertex AI Model Garden as claude-fable-5-1@20260902 (check the current model version string in the console). Enable the Anthropic Claude models API in your GCP project, request quota if needed, and call through the Vertex predict endpoint or the Gen AI SDK. Billing flows through your GCP invoice; retention and safety rules are Anthropic's, not Google's.
Support status
Supported in Vertex AI Model Garden, verified September 2, 2026. Google added the Claude Fable 5.1 model card alongside the September Anthropic release. Availability is region-specific — not every GCP region carries every Claude model on day one. Enterprise features (VPC Service Controls, CMEK, Cloud Audit Logs) wrap the call, but Anthropic's 30-day mandatory retention on Fable 5.1 still applies to the inference payload.
Setup
Five steps — quota and region are where teams stall.
- 1
Enable Vertex AI and the Claude models API
In Google Cloud Console, enable the Vertex AI API and the Anthropic Claude models partner API for your project. Model Garden listings appear only after both are active.
- 2
Confirm region availability
Open Model Garden and filter for claude-fable-5-1 in your target region (commonly us-east5 or europe-west1 for Claude models — verify in console). Deploying to an unsupported region returns a model-not-found error that looks like a typo.
- 3
Request quota if the default is zero
New projects often start with zero TPM quota for partner models. File a quota increase request in IAM & Admin → Quotas, filtering for Vertex AI Anthropic Claude. Approval can take one to three business days.
- 4
Call with the versioned model ID
Use the full model resource name from Model Garden — typically projects/undefined/locations/undefined/publishers/anthropic/models/claude-fable-5-1@20260902. Pass effort through the Anthropic-compatible parameters; omit temperature.
- 5
Wire observability and fallback
Enable Cloud Logging for predict calls. Configure client-side fallback for stop_reason refusal — Vertex forwards Anthropic refusals as successful responses, same as the direct API.
Gotchas
Enterprise wrappers do not rewrite Anthropic's model rules.
Default quota is often zero
Unlike direct Anthropic keys with default rate limits, Vertex partner-model quota starts at zero for many projects. Production deploys fail with RESOURCE_EXHAUSTED until quota is approved — plan ahead.
Region list changes
Claude models roll out region by region on Vertex. Hard-coding a region in infrastructure-as-code without checking Model Garden leads to drift when Google expands availability — or shrinks it during incidents.
VPC-SC does not remove 30-day retention
VPC Service Controls keep traffic inside your GCP perimeter, but Fable 5.1's mandatory 30-day retention is an Anthropic policy on the inference payload. VPC-SC is not a substitute for zero-data-retention — if ZDR is required, Fable 5.1 is not an option.
GCP billing ≠ Anthropic console visibility
Usage appears on your GCP invoice under Vertex AI partner model charges, not in the Anthropic console. Cross-charging teams need a billing export or labels — do not assume the Anthropic dashboard reflects Vertex-routed traffic.
Cost notes
Vertex list rates for Fable 5.1 mirror Anthropic published pricing ($10/$50 per million input/output as of September 2026) plus standard GCP egress where applicable. Committed use discounts on GCP compute do not apply to partner model inference. Prompt caching discounts apply when you send cache_control markers through the Vertex Anthropic-compatible endpoint.
Compare Vertex vs direct Anthropic billing
If you are already choosing between cloud fronts, our AWS Bedrock guide covers the same trade-offs from the other hyperscaler.