Google Vertex AI in Practice

Claude Model Versions on Vertex AI: Availability and Deprecation

A model you build on today will eventually be retired. On Vertex AI the lifecycle dates are set by Google, not Anthropic — which changes how you should plan upgrades.

Claude 3P 101 · Updated July 2026 · Unofficial guide

Claude models arrive on Vertex AI as partner models: Anthropic ships a new version, Google lists it in Model Garden, and you enable it per model card. As of July 2026 the roster runs from the current generation (Claude Fable 5, Opus 4.8, Sonnet 5) back through Opus 4.7/4.6/4.5/4.1/4, Sonnet 4.6/4.5/4, Haiku 4.5, and several Claude 3.x models. Recent launches have been effectively simultaneous across platforms — Claude Fable 5, for example, became generally available on the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud, and Microsoft Foundry on the same day, June 9, 2026. Availability on the way in is rarely the problem. The way out is what needs managing.

Two lifecycles, not one

Anthropic publishes deprecation and retirement dates for its first-party API. Google publishes its own lifecycle dates for the same models on Vertex — and they are not copies of each other. Google's Claude partner-model page states retirement floors such as: Claude Fable 5 retires "not sooner than June 8, 2027"; Opus 4.8 not sooner than May 28, 2027; Sonnet 5 not sooner than December 24, 2026. Meanwhile some models retired on the first-party API live on with Google: per Anthropic's pricing docs, Claude Opus 4 is retired except on Google Cloud, and Claude Sonnet 4 and Haiku 3.5 are retired except on Bedrock and Google Cloud. The practical readings: a Vertex deployment can sometimes keep an old model longer than a 1P deployment could, and you must track Google's dates, not just Anthropic's, because Google's floor dates ("not sooner than") are commitments about the earliest possible retirement, not exact end-of-life days.

Model (Vertex ID)Google retirement floor
claude-fable-5Not sooner than June 8, 2027
claude-opus-4-8Not sooner than May 28, 2027
claude-sonnet-5Not sooner than December 24, 2026

Version pinning: what the IDs mean

On Google Cloud, Claude model IDs are bare Anthropic-style names with no anthropic. prefix (that prefix is a Bedrock convention). Current-generation models use dateless IDs — claude-opus-4-8, claude-sonnet-5, claude-fable-5 — while some older models carry a version-date suffix in the @date form, such as claude-haiku-4-5@20251001 or claude-sonnet-4-5@20250929. Don't mistake dateless for "auto-upgrading": per Anthropic's models documentation, every Claude model ID is a pinned snapshot; starting with the 4.6 generation the IDs are simply dateless while remaining pinned. Your model string will not silently change behavior underneath you — behavior changes only when you change the string.

A version upgrade workflow that survives contact with production

  1. Inventory. Keep model IDs in configuration, not scattered through code, so "what runs where" is answerable in one query.
  2. Watch both notice channels. Google's Claude partner-model page for Vertex lifecycle dates; Anthropic's model overview and migration guide for deprecations and behavioral changes (e.g., claude-opus-4-1-20250805 is deprecated and retires August 5, 2026, with Opus 4.8 as the designated migration target).
  3. Read the migration guide before flipping the string. Upgrades across generations can be breaking — sampling parameters, extended-thinking configuration, and prefill behavior have all changed between recent generations — while some steps (Opus 4.7 → 4.8) are documented as fully backward compatible.
  4. Run old and new side by side. Evaluate the new version on your real prompts and traffic samples; see evaluation and testing.
  5. Cut over with a rollback path, keeping the old ID enabled until the new one has soaked.
Quota is one less thing: Claude models launched on Vertex after May 26, 2026 share model-lineage quota buckets (e.g. all Opus versions draw from anthropic-claude-opus per location), so adopting a new version within a lineage needs no new quota request. Note that changing the model string invalidates your prompt cache — the first request on the new model rewrites it.

Where to go next

See upgrading models for the cross-platform playbook, version pinning for the general policy question, and the Vertex model ID format guide for ID syntax details.

Sources