Amazon Bedrock offers two ways to pay for Claude. On-Demand is the default: you are charged per input and output token consumed, at list prices that match Anthropic's first-party rates. Provisioned Throughput is the alternative: fixed-cost hourly billing per model unit (MU), where each MU delivers a defined level of input tokens per minute and output tokens per minute. You are buying a throughput floor, not tokens — the meter runs whether or not traffic arrives.
How the model units pricing model works
A model unit is a capacity slice: so many input tokens per minute and so many output tokens per minute for a specific model. You purchase one or more MUs, and AWS bills you hourly per MU. Commitment options are no commitment, 1 month, or 6 months — the longer the commitment, the lower the hourly price. AWS does not publish a universal per-MU dollar figure that applies to every model and term in its general documentation; the concrete rates for your model, region, and term come from the Bedrock pricing page and your AWS conversation, so treat any specific MU price you hear second-hand as unverified until AWS quotes it.
Two structural notes matter for Claude specifically. First, Provisioned Throughput lives on the bedrock-runtime endpoint (the legacy InvokeModel/Converse surface); it does not apply to the newer bedrock-mantle endpoint that serves the current "Claude in Amazon Bedrock" Messages API. Second, cross-region inference profiles do not support Provisioned Throughput — provisioned capacity is tied to a region, which also means you keep the regional pricing and availability characteristics that come with that. Check which Claude models are currently offered for Provisioned Throughput in your region before planning around it.
When reserved capacity beats on-demand
The break-even logic is utilization. Compute what your sustained traffic costs on-demand, then compare it with the fixed hourly fee AWS quotes for enough MUs to carry that traffic. As a worked on-demand baseline using real list prices: a workload holding a steady 100,000 input and 20,000 output tokens per minute on Claude Opus 4.8 ($5/$25 per million tokens) costs $0.50 + $0.50 = $1.00 per minute on-demand — about $43,200 over a 30-day month. If your traffic actually runs at that level around the clock, a provisioned quote below that figure wins. If traffic runs twelve hours a day, your on-demand cost halves to roughly $21,600 while the provisioned meter keeps running — the quote now has to beat a much lower bar.
The capacity and latency angle
Cost is not the only reason enterprises buy reserved capacity. On-Demand traffic shares regional capacity and per-account quotas with everyone else; under throttling you see 429-style rejections and must retry. Provisioned Throughput gives you a dedicated throughput level that is yours by contract, which is primarily a predictability benefit: your P99 stops depending on shared-pool contention and quota headroom. AWS documentation frames it as guaranteed throughput rather than a specific latency SLA, so validate tail-latency behavior with your own load tests rather than assuming a number.
Also note the special case: custom models on Bedrock require Provisioned Throughput to serve at all, priced as their base model. For standard Claude models it is optional.
Opening the conversation with AWS
Committed-use economics on Bedrock are negotiated with AWS, not Anthropic. Come to your AWS account team with three numbers: your sustained tokens-per-minute (input and output separately, since MUs are defined on both), your monthly on-demand spend from Cost Explorer, and your growth forecast. Ask specifically: which Claude models and regions currently support Provisioned Throughput; what one MU delivers for your target model; and quotes for no-commitment, 1-month, and 6-month terms so you can price the flexibility you are giving up. A common pattern to discuss is a hybrid: provisioned capacity for the traffic floor, with overflow going to on-demand.
Where to go next
The decision framework with worked utilization math is in When Provisioned Throughput Pays Off. For the operational side of Bedrock capacity, see the Bedrock provisioned throughput setup guide and Bedrock quota types; for the broader negotiation picture, committed-use discounts.