Comparison

Bedrock Provisioned Throughput Pricing vs a Fixed-Price Plan (2026)

Back to BlogWritten by Published Nov 16, 2026
Bedrock Provisioned Throughput PricingAWS Bedrock Model Units CostBedrock Provisioned Throughput vs On-DemandBedrock Model UnitsFixed-Price Inference PlanGPU Cloud PricingCost Comparison
Bedrock Provisioned Throughput Pricing vs a Fixed-Price Plan (2026)

Amazon Bedrock's no-commitment Provisioned Throughput tier reads like a low-risk way to test reserved capacity before signing a 1-month or 6-month term. It mostly isn't. AWS restricts no-commitment purchases to custom models only, caps them at two Provisioned Throughputs per account, and for the base models most teams actually want, Claude, current-generation Llama, publishes no public rate at all. Bedrock Provisioned Throughput pricing on AWS's own docs covers the mechanics (Model Units, commitment terms, the purchase API) and skips the one number that decides whether any of it is worth signing: what a Model Unit actually costs for your model.

TL;DR: Bedrock Provisioned Throughput Pricing

Bedrock Provisioned Throughput bills per Model Unit per hour, with the rate set by model, MU count, and commitment term; AWS publishes it for only a few older models.

  • Published rate: Cohere Command runs $49.50/hr per Model Unit with no commitment, down to $23.77/hr on a 6-month term, as of 16 Nov 2026.
  • No-commitment cap: It's limited to custom models, capped at 2 per AWS account with 1 Model Unit each.
  • Alternative: Spheron's fixed-price plans state capacity as concurrent requests and a monthly token ceiling, with a trial key before any term. Compare GPU pricing.

What a Model Unit Actually Buys You

A Model Unit is Bedrock's throughput currency, and it isn't a flat number across models. Each Model Unit specifies the number of input tokens it can process and the number of output tokens it can generate, each counted across all requests within a one-minute span. A model with a higher per-token cost or a larger context window needs more throughput per request, so the same Model Unit buys a different real-world request rate depending on which model sits behind it.

That's also where the pricing trail runs cold. AWS directs buyers to contact their AWS account manager for the throughput number and the rate. You can read exactly how MUs are defined and purchased without ever learning what one costs for Claude or Nova.

There's a second wrinkle that makes Provisioned Throughput unavoidable for a specific group of buyers: if you customize a model inside Bedrock, you have to purchase Provisioned Throughput to use it at all. Customized models carry no on-demand tier. That's different from the on-demand side of Bedrock pricing, which our AWS Bedrock pricing breakdown covers in full; Provisioned Throughput is the layer that sits underneath it for fine-tuned and high-volume base-model deployments alike.

The Three Commitment Terms Behind Bedrock Provisioned Throughput Pricing

Every Provisioned Throughput purchase picks one of three terms, and the term is what moves the hourly rate. No commitment bills hourly with no lock-in and the highest per-MU rate. A 1-month term drops the rate in exchange for a month you can't exit early. A 6-month term drops it further still, the steepest discount AWS publishes for any model it shows pricing on at all.

The mechanics live in the CreateProvisionedModelThroughput API, and they're stricter than the docs prose suggests. The commitmentDuration field only accepts two values, OneMonth or SixMonths. And by default, an AWS account has zero Model Units available for purchasing commitment-based Provisioned Throughput at all. Before you can call that API successfully, with or without a commitment, you first have to file a service-limit-increase case in AWS Support Center to have Model Units distributed to your account. There's no self-serve path from "I want Provisioned Throughput" to a running deployment; a support ticket sits in front of the purchase API either way.

That three-way term structure isn't unique to Bedrock. It's the same reserved-capacity shape as Azure OpenAI's Provisioned Throughput Units, hourly billing per unit with reservation discounts for longer terms, and the same shape our serverless vs dedicated vs reserved pricing framework uses to sort every inference billing model on the market into one of three buckets. Bedrock's version just adds the support-ticket step on top.

Why No-Commitment Provisioned Throughput Is Capped at Custom Models Only

The obvious move for a buyer who doesn't want a 1-month or 6-month commitment is to just buy Provisioned Throughput with no commitment and pay the hourly premium for the flexibility. AWS doesn't let most buyers make that move. No-commitment Provisioned Throughput isn't available for Bedrock's base models at all, Claude, Nova, Llama, Mistral, none of them. It's available only for custom models, and even there it's capped at a maximum of 2 Provisioned Throughputs per account, with each one getting a quota of just 1 Model Unit.

In practice, that means the only no-commitment path through Provisioned Throughput runs through fine-tuning a model first, then buying two single-MU reservations against it. If your model is a stock base model and you don't want to sign a term, there's no no-commitment door to walk through; you go straight to 1-month or 6-month, on a rate AWS mostly won't show you in advance.

The restriction tightens further outside commercial regions. In the AWS GovCloud (US-West) Region, Provisioned Throughput is supported only for custom models purchased with no commitment; the 1-month and 6-month terms aren't offered there at all.

The Published Rates AWS Does Show (and the Models It Doesn't)

AWS does publish exact per-Model-Unit hourly rates, just not for the models most teams are deploying in 2026. Here's the full public rate card, by commitment term:

ModelNo commitment1-month6-month
Cohere Command$49.50/hr$39.60/hr$23.77/hr
Cohere Command-Light$8.56/hr$6.85/hr$4.11/hr
Cohere Embed 3 (English/Multilingual)$7.12/hr$6.76/hr$6.41/hr
Meta Llama 2 13B/70B (Pretrained/Chat)Not published$21.18/hr$13.08/hr

That's the entire public list. For the models not shown, AWS's pricing page states: "For Provisioned Throughput pricing, please reach out to your account team." Amazon Nova and current-generation Llama carry the same gap, no rate on the public page, a conversation instead. And Llama 2's Pretrained models specifically are only available in Bedrock via Provisioned Throughput after customization, so even the one Llama row in that table isn't a model you can reserve straight off the shelf.

The pattern across the table is worth naming plainly: the models AWS prices in public are older, lower-demand, and increasingly peripheral to what teams are actually deploying. The models driving most Bedrock Provisioned Throughput spending in 2026, current Claude and Nova generations, sit behind an account-manager conversation with no public anchor price at all.

Third-party rates above are published figures from AWS's own pricing and documentation pages as of 16 Nov 2026 and may have changed since. Check AWS's Bedrock pricing page for current published rates, and your AWS account team for any model not listed there.

The Under-Utilization Trap: Committing Before 90 Days of Real Volume

Sizing a 1-month or 6-month Provisioned Throughput commitment means guessing two things at once: the throughput a Model Unit delivers for your model, which AWS won't publish, and the traffic volume you'll actually push through it, which you often don't have real data on yet if you're moving off on-demand or off a different provider entirely. Get either number wrong and you're paying an hourly reservation rate for capacity that sits idle, with no per-token fallback inside that reservation.

The general shape of that risk isn't specific to Bedrock. Across fixed-capacity inference pricing, a fixed-price plan beats pay-as-you-go once traffic fills roughly 60-80% of the monthly allowance; below that, metering wins even at a higher per-token rate. Bedrock's Provisioned Throughput sits on the expensive side of that line by design: there's no metered fallback once you've committed, and no partial-month exit on a 1-month or 6-month term. Together AI's Dedicated Endpoints carry the same structural risk, a GPU reserved for you bills whether or not you send it traffic, which is the reason committing before you've watched a real quarter of production volume is the single most common way reserved inference capacity turns into a loss rather than a discount.

Where a Fixed-Price Counted-Token Plan Removes the Commitment Guess

Spheron's managed inference plans sidestep the Model Unit conversion problem by stating capacity in terms a buyer can size against directly. Starter covers 5 billion counted tokens a month. Growth covers 20 billion counted tokens a month, on a 3-month term. Both run on reserved GPUs, never spot, the same reliability premise a Model Unit reservation is selling, and a capped trial key lets you run real traffic before signing anything, a path no-commitment Provisioned Throughput doesn't offer unless you've already customized a model.

Token counting on these plans works on a weighted basis rather than a raw count: every output token counts 4x against the monthly allowance, every fresh input token counts 1x, and cached input counts 0.25x. That weighting is a different sizing exercise than converting requests-per-minute into Model Units, but it's one you can run against your own traffic logs without a support ticket or an account-manager call.

This isn't a drop-in replacement for Bedrock, and it's worth being direct about where it isn't. A team already committed to a Bedrock-exclusive model family may have no path off Model Units through this plan. Pricing is quoted per request, not published, the same gap as Bedrock's account-manager-only MU rates; neither side hands you a number without a conversation. And Growth's 3-month term is itself a commitment decision, it doesn't remove sizing risk entirely, it just states the guarantee in concurrency and tokens instead of Model Units and tokens-per-minute. What it does remove is the two-step guess Bedrock forces on most buyers: figuring out what an MU delivers, then figuring out how many you need, before AWS will even tell you the price.

Request a Plan

If guessing a Model Unit's throughput before AWS will quote you a rate is the part of Provisioned Throughput sizing you'd rather skip, a fixed-price plan states the same reserved-GPU guarantee in concurrency and tokens instead.

Request a fixed-price inference plan →

FAQ / 04

Frequently Asked Questions

It depends on the model, the number of Model Units, and the commitment term, and AWS only publishes that rate for a handful of older models. Cohere Command runs $49.50/hr per Model Unit with no commitment, $39.60/hr on a 1-month term, and $23.77/hr on a 6-month term. For Claude, Amazon Nova, and current-generation Llama, AWS's pricing page tells you to reach out to your account team instead of showing a number.

A fixed hourly reservation of model capacity, measured in Model Units, that you pay for whether or not you send any requests. One Model Unit sets the number of input tokens it can process and output tokens it can generate, across all requests, within a one-minute window. It's the opposite of on-demand Bedrock, which bills per token with no reservation and no idle cost.

Only for custom models. No-commitment Provisioned Throughput isn't available for Bedrock's base models at all; it's capped at a maximum of 2 Provisioned Throughputs per account, with each one getting a quota of 1 Model Unit. If you want no-commitment capacity on a base model like Claude or Llama, that path doesn't exist, only the 1-month and 6-month terms do.

AWS's own Bedrock pricing page states plainly, for Anthropic models, to reach out to your account team for Provisioned Throughput pricing. That's a deliberate gap, not a documentation oversight: the handful of models AWS does price publicly (Cohere Command, Cohere Command-Light, Cohere Embed, Meta Llama 2) are older and lower-demand, while current-generation Claude and Llama capacity gets priced per account.

// Try It Yourself

Try It on Real GPUs

The GPUs behind these guides are the ones you can rent here: H100s, H200s, B200s, and more, billed per minute after a 20-minute minimum runtime, with no contracts. Pick one and you are live in under two minutes.

View Pricing
→
Or talk to our team
Spheron GPU instance going live and linking data centers across the globe