Historically, each Claude model version on Google Vertex AI carried its own quota: so much capacity for one Opus version, separately negotiated capacity for the next. Google has changed that model. Per its quota documentation, Claude models launched after May 26, 2026 use shared model-lineage quotas: one quota bucket per lineage per location. All Opus versions covered by the scheme draw from a single anthropic-claude-opus pool on the global endpoint, rather than each version having its own allocation.
A "lineage" here is just the model family. Google's documented base_model quota dimensions are anthropic-claude-opus, anthropic-claude-sonnet, anthropic-claude-haiku, and anthropic-claude-fable. The quota metrics themselves are family-scoped too — for the global endpoint: global_online_prediction_requests_per_base_model, global_online_prediction_input_tokens_per_minute_per_base_model, and global_online_prediction_output_tokens_per_minute_per_base_model.
The upgrade benefit
The headline advantage is operational. When Anthropic ships a new Opus version and Google makes it available, it lands inside the existing Opus bucket — Google states that adding a new version of a lineage requires no new quota request. Teams that remember pre-2026 model launches (file the increase request early, wait for approval, hope capacity arrives before your rollout date) will appreciate what this removes from an upgrade checklist. You can canary the new version and shift traffic across at your own pace, all inside quota you already hold.
The flip side: versions compete
A shared bucket cuts both ways. During a migration, the old and new Opus versions are not independent workloads from the quota system's perspective — they are one workload. If you run a dual-stack period where both versions serve production traffic at full volume, their combined consumption draws down one pool, and a spike on either can throttle both. Plan migration overlap the way you'd plan any traffic increase within a fixed allocation: the bucket needs headroom for the sum, not the max, of the two versions while they coexist.
There's a subtler planning wrinkle, too. Newer Claude models (Opus 4.7 and later, Sonnet 5, Fable 5) use a tokenizer that produces roughly 30% more tokens for the same text than earlier generations, per Anthropic's documentation. Migrating within a lineage can therefore raise your token-per-minute consumption against the shared bucket even at identical request volume — worth folding into the forecast before flipping traffic.
What the scheme does and doesn't cover
Three boundaries keep the mental model honest:
Launch-date cutoff. Only models launched after May 26, 2026 use lineage quotas. Models launched before that date keep per-model quotas, measured in queries per minute (QPM) and tokens per minute (TPM, input + output combined), varying by endpoint type, with regional metric names like online_prediction_requests_per_base_model. A mixed estate can be under both schemes at once.
Location boundaries. Lineage buckets are per location, and Google documents that the global-endpoint quota and each multi-region-endpoint quota are independent buckets — usage on one doesn't consume the other. An Opus pool on the global endpoint says nothing about your Opus capacity on the us multi-region endpoint. See multi-region quota isolation for how to use that deliberately.
Lineage boundaries. Opus and Sonnet never share a bucket. Moving load from one family to another — say, offloading simple requests to Haiku — genuinely frees Opus quota, which makes model choice a quota-pool decision, not just a cost decision.
Watching a shared bucket
Because several versions share one meter, per-version dashboards won't explain throttling on their own; you need the lineage-level quota metrics above, filtered by base_model. One caution when reading them: Google documents that token usage shown on the console Quota page can be inaccurate due to Anthropic's token estimation and refund system — a known reporting gap — so use the token-counting API or token_count metrics in Metrics Explorer for numbers you intend to plan against.
Where to go next
See Vertex quota types for the full taxonomy, requesting a Vertex quota increase for the process, and provisioned throughput on Vertex when pay-as-you-go quota stops being enough.