Claude Haiku Vertex AI Shutdown: Haiku Is Gone, Opus Is Next — Audit Your Partner-Model Routing Now

Vertex AI silently closed one door yesterday and will close another in 26 days. Claude 3.5 Haiku hit its shutdown date on July 5, 2026 — any API call to that model ID on Vertex now returns an error. Claude 3 Opus was deprecated on June 30 and exits on August 1. If your routing policy targets either of these models on Vertex AI, the decision has already been made for you.

The operational challenge is not just "update a model ID." It is that Vertex AI's partner-model deprecation lifecycle runs on a different clock than the direct Anthropic API — and teams that route across both surfaces may have migrated on one track without realizing the other was still live and expiring on its own schedule.

What happened

Google's Vertex AI (now rebranded to Gemini Enterprise Agent Platform) hosts Anthropic Claude models under a managed API with its own deprecation timeline, separate from the direct Anthropic API at api.anthropic.com.

The current status on Vertex AI is as follows:

• Claude 3.5 Haiku — Deprecated January 5, 2026. Shut down July 5, 2026. Requests to this model ID now return errors.
• Claude 3 Opus — Deprecated June 30, 2026. Shutdown: August 1, 2026. Twenty-six days remain.
• Claude Sonnet 5 — Available, retirement not sooner than December 24, 2026.
• Claude Fable 5 — Available, retirement not sooner than June 8, 2027.

For reference: on the direct Anthropic API, Claude Haiku 3.5 was retired February 19, 2026 — nearly five months before the Vertex AI shutdown. Teams that checked the direct API docs may have seen "retired" and assumed the same applied to Vertex. It did not.

Why it matters for AI engineering teams

Two-track deprecation is an operational trap. Anthropic and Google publish separate deprecation schedules. A model that is retired on api.anthropic.com may still be available on Vertex AI, and vice versa. If your monitoring, fallback logic, or hardcoded model IDs only follow one provider's docs, you miss the other track.

For teams that route to Claude through Vertex AI — using Google Cloud credentials, VPC Service Controls, or provisioned throughput — this creates an immediate exposure:

1. Haiku (already gone): Any workload pointing to claude-3-5-haiku@20241022 on Vertex is now returning HTTP errors. If you have no fallback and no alerting on model-level 4xx rates, this may be silent for hours.
2. Opus (26 days): Claude 3 Opus via Vertex AI will follow on August 1. Opus on the direct API was retired alongside Claude 4 on June 15, 2026. Teams who migrated direct-API Opus traffic may have left Vertex-side Opus routing untouched.

The split is also a billing consideration: provisioned throughput capacity purchased for Vertex AI partner models does not automatically transfer to a new model, and provisioned slots for deprecated models may stop being serviceable even before the official shutdown date.

The router/operator angle

Three decisions every routing team must make before August 1:

1. Inventory your Vertex AI model IDs. Pull your active routing configurations and filter for any claude-3-5-haiku*, claude-3-opus*, or claude-3-* model identifiers that point to Vertex AI endpoints. Do not rely on traffic being zero as proof they are unused — low-volume fallback routes are easy to miss.

2. Map the Vertex-to-Anthropic ID gap. Vertex AI model IDs follow the format claude-sonnet-5@*, claude-fable-5@*, etc. The direct Anthropic API uses claude-sonnet-5, claude-fable-5. If your routing layer normalizes model IDs across providers, verify the mapping is current. Outdated aliases that pointed to Haiku or Opus may now need to point to Sonnet 5 or Fable 5.

3. Set up provider-aware fallback. A model removal on one provider does not mean the same model is gone everywhere. If you route to Vertex AI for compliance (VPC Service Controls, data residency) but the model is gone, your fallback should route to another Vertex-available Claude model — not necessarily to the direct Anthropic API, which may not satisfy the same compliance constraints.

What to watch

• Claude 3 Opus on Vertex AI exits August 1. This is the immediate deadline.
• The Vertex AI partner-model page (cloud.google.com/vertex-ai/generative-ai/docs/partner-models/use-claude) now lists only Claude Sonnet 5 and Claude Fable 5 as available models. Use this as the ground truth for what is currently live.
• Provisioned throughput holders: check with your Google Cloud account team about what happens to any committed throughput for a model that is shutting down.
• On the direct Anthropic API, Claude 3 Opus on Bedrock (AWS) follows yet another schedule — check the Bedrock partner model deprecation docs separately.

If you route Claude workloads through multiple providers, now is the time to build a scheduled check that compares your active routing table against each provider's live model catalog. Routing to a model ID that no longer exists does not fail gracefully — it fails immediately.