Claude 3 Haiku stops serving requests on Google Cloud’s Vertex AI today. Google marked the model deprecated on 23 February 2026 and set the shutdown for six months later, 23 August — the date on its partner-model deprecation page, and the date that lands this morning.
The wrinkle is that Anthropic already killed the same model on its own API in April. Developers calling claude-3-haiku-20240307 direct got their notice in February and lost the endpoint on 19 April. Anyone routing the identical model string through Vertex kept running it for another four months, entirely legitimately, on a schedule Google set independently.
So the same model has had two different funerals, four months apart, depending on which door you knocked on. That is not a bug in anyone’s process. It is how the partner-model arrangement works — and it is the part most teams have not internalised.
Three calendars, not one
Amazon Bedrock and Google Vertex AI each publish and manage their own model lifecycle schedules for the Claude family. Those schedules can differ from Anthropic’s, and from each other’s. A shop running Anthropic direct for production and Vertex for a regulated workload is watching three deprecation calendars, and only one of them sends the email everyone actually reads.
The failure mode is not dramatic. Nothing crashes at 2am. A batch job that has run untouched for eighteen months starts returning errors on a model string a colleague migrated off in April, and the person who wrote it has since moved teams.
Our take: The multi-cloud pitch was resilience — if one provider goes down, route to another. What nobody priced in was that multi-cloud also multiplies your deprecation surface. Every additional route to the same model is another calendar that can retire it out from under you on a date you did not choose.
The replacement costs four times as much
Claude 3 Haiku ran at $0.25 per million input tokens and $1.25 per million output. Claude Haiku 4.5, the current small model in the family, is $1.00 and $5.00. That is a straight 4x on both sides of the ledger.
For a chat feature, nobody notices. For the workloads Haiku 3 was actually chosen for — classification, extraction, routing, tagging, the unglamorous high-volume plumbing where the whole point was that tokens were nearly free — a 4x turns a rounding error into a line item finance asks about. Batch pricing halves it. It does not erase it.
Which is why the honest migration question is not “which Claude replaces this,” but “does this job still need a frontier-lab model at all.” Plenty of the work Haiku 3 absorbed in 2024 is now handled by cheaper small models from Google and OpenAI, or by open weights running on hardware you already rent. A forced migration is the cheapest moment you will ever get to re-run that comparison, because you are rewriting the call site regardless.
What to watch
- Your own model strings. Grep the codebase for hardcoded dated identifiers. The ones buried in config files and notebooks are the ones that bite.
- Per-cloud deprecation pages, not just the lab’s. Vertex and Bedrock publish separately. Subscribe to both if you route to both.
- The next tier up. Retirement waves move through a model family over time. Small models go first, but the schedule does not stop there.
- Whether cheap-model competition holds. Haiku 3 to Haiku 4.5 was a 4x. If that becomes the pattern at the bottom of every model family, high-volume AI plumbing gets repriced across the industry, not just at one lab.
The switch flips today. If you find out from a dashboard rather than a support ticket, you got lucky.