Claude 3 Haiku on Vertex AI is now shut down: here's what to do if your app is broken
Anthropic's Claude 3 Haiku on Vertex AI reached end-of-life on 23 August 2026. Any app still calling it will now receive errors.
If you have an application calling claude-3-haiku@20240307 on Google Cloud’s Vertex AI, it stopped working on 23 August 2026. That is the permanent shutdown date for Claude 3 Haiku on Vertex AI, and any API call to that model ID now returns an error.
This is not a soft deprecation or a warning phase. The model is gone from the platform.
What happened and when
Anthropic’s Claude 3 Haiku was one of the faster, lighter models in the Claude 3 family, designed for high-volume, low-latency workloads that needed text, code, and image inputs processed quickly. On Vertex AI it ran under the model ID claude-3-haiku@20240307, with a 200,000-token input window and up to 8,000 tokens of output.
Google marked it as deprecated on 23 February 2026, giving teams six months to migrate before the 23 August 2026 shutdown. That window has now closed.
One detail worth understanding: this timeline is specific to Vertex AI. On Anthropic’s direct API at api.anthropic.com, Claude 3 Haiku was retired earlier, on 20 April 2026, after a deprecation notice in February. Vertex AI’s shutdown came approximately four months later, which may have created a false sense of security for teams that migrated away from the direct API and assumed the Vertex AI endpoint was on the same schedule. It was not.
Why this catches teams off guard
Google Cloud hosts Anthropic’s models through a managed Model as a Service arrangement, and it runs its own deprecation calendar independently of Anthropic’s. The two clocks do not align, and there is no automatic notification that bridges the two platforms.
If your team migrated from api.anthropic.com back in April and ticked the box, but had a separate integration routed through Vertex AI, that second integration has been broken since 23 August 2026. It is a common enough failure mode that it is worth checking explicitly, especially in organisations where infrastructure is split across teams or where Vertex AI integrations were set up as part of a broader Google Cloud deployment rather than as a standalone AI project.
Claude 3 Haiku is described in Google’s documentation as the last Claude 3-generation model to be shut down on the platform, so this closes out the entire Claude 3 family on Vertex AI.
What to do right now
If you are seeing errors, the fix is straightforward in most cases: update the model ID in your code or configuration. You do not need to change your authentication setup or your anthropic_version header, which stays as vertex-2023-10-16.
The question is which model to move to. A few options worth considering:
Claude Haiku 4.5 on Vertex AI (claude-haiku-4-5@20251001) is the direct successor. It is faster and more capable than Claude 3 Haiku, and pricing on batch workloads comes in at around $0.50 per million input tokens and $2.50 per million output tokens. That is roughly double the price of Claude 3 Haiku, but for most workloads the capability improvement means you may be doing more work per call, which can offset some of the cost difference.
Gemini 2.5 Flash-Lite is worth considering if you want to stay within Google Cloud and keep costs down. At approximately $0.10 per million input tokens and $0.40 per million output tokens, it is significantly cheaper, and it is a native Vertex AI model with no cross-platform dependency to manage.
The right choice depends on how much your application relies on Claude-specific behaviour. If you have prompt logic tuned to Claude’s response style, Haiku 4.5 will be the lower-friction migration. If you are doing more generic classification, summarisation, or extraction tasks, Flash-Lite is worth testing.
The broader pattern to watch
Claude 3 Haiku is not the only model going through this cycle. Several other Claude models on Vertex AI have their own upcoming shutdown dates:
- Claude 3.5 Haiku on Vertex AI was deprecated on 5 January 2026 and shuts down on 5 July 2026.
- Claude 3.7 Sonnet on Vertex AI was deprecated on 11 November 2025 and shut down on 11 May 2026.
- Claude 3.5 Sonnet on Vertex AI was deprecated on 20 August 2025 and shut down on 19 February 2026.
If you have applications on Vertex AI calling any of these model IDs, check the Google Cloud partner model deprecations page now rather than waiting for an error report to surface in production.
The practical lesson here is to treat Vertex AI and api.anthropic.com as two separate platforms with two separate deprecation schedules, and to check both independently whenever you are planning a migration. Building a note into your infrastructure review process to cross-reference Anthropic’s model deprecations page alongside Google’s documentation will save a painful incident in the future.