Claude on Vertex AI defaults to pay-as-you-go: per-token billing at list prices that match Anthropic's first-party rates (as of July 2026, Claude Opus 4.8 at $5/$25 per million input/output tokens, Sonnet 5 at $3/$15 standard with a $2/$10 promotion through August 31, 2026, Haiku 4.5 at $1/$5 on the global endpoint). When enterprises want price predictability or guaranteed capacity, the documented structure is Provisioned Throughput — plus whatever negotiated terms your broader Google Cloud agreement carries.
What Provisioned Throughput on Vertex actually is
Google describes Provisioned Throughput as "a fixed-cost, fixed-term subscription available in several term-lengths that reserves throughput for supported generative AI models." For partner models like Claude, it "reserves throughput capacity for your models for a fixed fee," and provisioned requests "are prioritized over the standard pay-as-you-go requests." You specify the model and locations; to subscribe, Google's documentation says to contact sales — there is no self-serve checkout for partner-model commitments.
Three constraints shape whether it fits your architecture:
| Constraint | Detail |
|---|---|
| Regional endpoints only | Provisioned Throughput is not supported on the global endpoint or the us/eu multi-region endpoints. Global and multi-region are pay-as-you-go only. |
| Model coverage lags the newest models | Google's pricing page lists Claude models supported for Provisioned Throughput as Haiku 4.5, Sonnet 4.6, Sonnet 4.5, Sonnet 4, Opus 4.6, Opus 4.5, Opus 4.1, and Opus 4 — notably not (as of this writing) Sonnet 5, Opus 4.8, or Fable 5. Confirm current coverage with Google. |
| Fixed term, fixed fee | Several term lengths exist; the fee is committed regardless of whether you use the capacity. Utilization is your risk. |
Note an interaction with pricing you already pay: regional and multi-region endpoints carry a 10% premium over global endpoints for Claude Sonnet 4.5 and later models. Since Provisioned Throughput requires regional endpoints, part of any commitment discussion is comparing against a global-endpoint pay-as-you-go baseline, not just the regional one.
Beyond Provisioned Throughput: the account-level lever
Committed-use discounts on cloud marketplaces are negotiated with the cloud provider, not with Anthropic. If your organization has an enterprise agreement or committed-spend contract with Google Cloud, Claude usage on Vertex is Google Cloud spend, and how it counts toward such commitments is a contract question for your account team. Nothing in the public Claude-on-Vertex documentation spells out custom discount mechanics, so do not assume specific percentages — ask.
How to approach the negotiation
Walk in with measured data rather than ambitions. Useful preparation:
1. Establish the pay-as-you-go baseline. Pull your actual Claude token consumption and spend from Cloud Billing, split by model and endpoint type. If you are not yet at scale, model it from list prices — e.g., a steady 50,000 input + 10,000 output tokens per minute on Opus 4.8 is $0.25 + $0.25 = $0.50/minute, roughly $21,600 over a 30-day month on-demand.
2. Decide what you are buying. If the goal is price predictability, the fixed fee versus that baseline is the whole conversation. If the goal is capacity assurance — avoiding quota contention during peaks — the prioritization of provisioned requests over pay-as-you-go is the value, and you may accept break-even economics for it. Vertex quotas for newer Claude models are shared per model lineage per location, so also ask whether a quota increase alone solves your problem more cheaply.
3. Ask the specific questions. Which Claude models and regions can be covered today? What throughput does one subscription unit deliver for your target model? What term lengths are offered and how does price vary by term? What happens to overflow traffic above the reserved level? Can the commitment be re-pointed if you migrate models mid-term — given that model coverage today skews to the 4.x generation, migration flexibility matters more than usual.
Where to go next
The general decision framework is in When Provisioned Throughput Pays Off. For Vertex specifics, see the Vertex provisioned throughput guide, Vertex quota types, and global vs regional endpoints.