Both AWS and Google Cloud offer a reserved-capacity alternative to on-demand Claude pricing. On Amazon Bedrock, Provisioned Throughput bills a fixed hourly rate per model unit (a defined level of input and output tokens per minute), with no-commitment, 1-month, and 6-month options — longer commitments, lower hourly price. On Vertex AI, Provisioned Throughput is a fixed-cost, fixed-term subscription that reserves throughput, available for partner models via Google sales and only on regional endpoints. Neither provider publishes a universal reserved dollar rate in the documentation we can cite, so the honest framework is: compute your on-demand baseline precisely, then judge any quote against it.
Step 1: price your traffic on-demand
On-demand cost is deterministic. Using Claude Opus 4.8 list pricing ($5 per million input tokens, $25 per million output tokens), a workload sustaining 100,000 input and 20,000 output tokens per minute costs:
| Component | Math | Cost |
|---|---|---|
| Input | 0.1M tokens/min × $5/M | $0.50/min |
| Output | 0.02M tokens/min × $25/M | $0.50/min |
| Total, 24/7 | $1.00/min × 60 × 24 × 30 | ≈ $43,200/month |
That $43,200 is the ceiling a reserved quote must beat — if the traffic truly runs around the clock.
Step 2: apply your real utilization
Reserved capacity bills continuously; on-demand bills only when tokens flow. If the same workload actually runs at full rate twelve hours a day and idles otherwise, its on-demand cost is about $21,600 a month — but the reserved meter still runs 24/7. In general, if your traffic occupies the reserved capacity a fraction U of the time, a reserved quote must be cheaper than on-demand price × U to save money. At 50% utilization, the reservation must effectively be half the on-demand rate just to break even; at 90%+ utilization, even a modest discount wins. This is why flat, machine-generated traffic (continuous document pipelines, round-the-clock global products) suits reservations, while business-hours or bursty human traffic rarely does.
Step 3: subtract the optimizations you'd otherwise use
Compare against your optimized on-demand cost, not the naive one. Prompt caching bills repeated prefixes at 0.1x input price; batch-eligible work runs at a 50% discount (Anthropic's Message Batches API on first-party and Claude Platform on AWS; AWS and Google Cloud's own batch inference mechanisms on Bedrock and Vertex). If 40% of the example workload could move to batch, the true comparison baseline drops accordingly. A reservation sized to unoptimized traffic locks in your inefficiency.
Step 4: price the non-cost benefits honestly
Sometimes reserved capacity is worth buying at break-even or worse, because you are buying assurance rather than discount:
Throughput assurance. On-demand traffic lives under shared quotas — Bedrock enforces per-model token and request quotas, and Vertex applies per-lineage quotas. Reserved capacity is a contractual floor that quota contention cannot take away.
Priority under load. Google documents that provisioned requests on Vertex "are prioritized over the standard pay-as-you-go requests." Fewer queued or rejected requests at peak translates into steadier tail latency, though neither provider's documentation promises a specific latency number — measure it in your own load tests.
Budget fixity. A fixed fee is a feature for finance even when it costs slightly more than metered spend would have.
The decision framework in one pass
bedrock-runtime surface, not bedrock-mantle, and not via inference profiles); and (4) the quoted fee beats on-demand × utilization after caching and batch optimizations. Otherwise, stay on-demand and buy a quota increase instead — it's free.A pragmatic hybrid deserves the last word: reserve the floor, overflow to on-demand. Size the reservation to the traffic level you sustain even in your quietest hours, and let metered pricing absorb everything above it. That keeps reservation utilization near 100% — the only regime where the math reliably works in your favor.
Where to go next
Platform specifics and negotiation prompts are in Provisioned Throughput on Amazon Bedrock and Committed Use on Google Vertex AI. The broader trade-off is framed in on-demand vs reserved, and smoothing spiky traffic first is covered in peak load smoothing.