Compare LLM capacity contracts and billing
Eight throughput and hardware offers compared by billing unit, commitment, and overflow, with retired Anthropic Priority Tier treated separately.
Provisioned throughput, dedicated GPU deployments, and monthly output allocations all let a team purchase inference resources ahead of use. They charge for different things. Compare the unit, billing meter, commitment, and overflow policy before comparing prices; a GPU-hour and a token-throughput unit do not represent the same workload.
The table uses official documentation snapshots from September 16, 2026, with Together checked September 29 and Standard Thinking October 1. Azure’s capacity and reservation distinction was rechecked October 5. Other prices and terms have not been refreshed to that date.
Throughput offers and dedicated deployments
Dedicated hardware deployments are included as an alternative purchasing type; they are not all reserved-throughput contracts. “None” means no minimum term documented for the listed on-demand option, not a promise of immediate capacity.
| Offering | Unit | Billing meter | Minimum term | Price visibility | Overflow |
|---|---|---|---|---|---|
| Azure provisioned throughput | PTU | Hourly | None | Partial | Configured spillover |
| Bedrock Provisioned Throughput | Model Unit | Hourly | Model-dependent | Sales quote | Check configuration |
| Vertex AI Provisioned Throughput | GSU | Hourly rate | 1 week | Public | Pay as you go |
| OpenAI Scale Tier | Token unit | Daily | 30 days | Public | Pay as you go |
| Together PTUs / hardware | PTU / GPU | Minute / hour | Check offer | Partial | Check offer |
| Fireworks on-demand deployment | GPU | Per second | None | Public | Hardware / scaling |
| Baseten dedicated deployment | GPU instance | Per minute | None | Public | Configured scaling |
| Standard Thinking Product Plan | Output tokens/s | Monthly + blocks | 1 month | Indicative | Extra capacity options |
Azure. A provisioned deployment holds processing capacity and is billed for allocated PTUs regardless of token consumption. Optional one-month or one-year Azure Reservations discount that billing meter; the reservation itself does not secure deployment capacity, and configured spillover is limited to supported Azure OpenAI models.
Bedrock. Model Units specify input and output tokens per minute across requests; hourly pricing and available commitments depend on the model. The documentation directs buyers to an account manager for unit specifications, and committed purchases cannot be deleted before the term ends.
Vertex AI. The September snapshot lists global-endpoint rates from $7.14 per GSU-hour for one week to about $2.74 for one year. Orders cannot be cancelled mid-term; model-specific burndown rules determine consumption, including cached input, and overages can use pay-as-you-go billing.
OpenAI. Scale Tier purchases throughput for a model snapshot with a 30-day minimum and enterprise access requirements. The September snapshot lists GPT‑5.5 at 50,000 TPM for $750 per day; combined bundles weight output differently from input, so this is not 50,000 output tokens per minute.
Together. PTUs specify model-dependent input, cached-input, and output rates; the model rows were unreadable in the recorded check, so no PTU price is quoted here. The September 29 hardware snapshot lists HGX H100 at $5.49 per GPU-hour and HGX B200 at $8.99, with reserved hardware quoted through sales.
Fireworks. On-demand deployments charge per GPU second and exclude startup time. The September snapshot’s “from Sep 1” column lists H100 80 GB at $8 per hour; the hardware rate establishes no token entitlement.
Baseten. Dedicated deployments charge per minute, with no idle charge under the documented rules. The September snapshot lists H100 80 GiB at $0.10833 per minute; configured scaling and measured serving performance determine the workload it can support.
Standard Thinking. Our Product Plan example uses an indicative $1,200 monthly model unit and $5 per extra unit per 2‑hour block for authorized auto scaling. Free shared-pool extras are not reserved, auto scaling depends on availability, and prices and service conditions are confirmed before paid use.
Historical: Anthropic Priority Tier
Anthropic’s September documentation states that Priority Tier capacity commitments are no longer purchasable. The retired offer committed input and output tokens per minute for a named model version, with standard-tier fallback above the commitment; it is historical context rather than another current buying option.
Make the units comparable
For throughput contracts, apply the provider’s input/output weighting, cache rules, model version, and measurement window to your actual requests. Converting tokens/minute to tokens/second only changes the time unit; it does not reconcile different definitions.
For GPU deployments, measure the chosen model and serving configuration before estimating answers per hour. Hardware type and rental price alone cannot supply that conversion. Include startup, idle behavior, replica count, and scaling policies in the cost scenario.
Then confirm the term, renewal, exit rights, regional availability, and overflow behavior. A published rate is useful for budgeting, but it does not establish that the allocation is available or that your latency target is covered. The inference buying guide explains how to compare the complete cost of a baseline and its peaks.
Sources and verification
Sources used for this resource. Verification dates describe our source checks, not provider effective dates.
Azure AI Foundry provisioned throughput documentation
Provisioned deployments hold model processing capacity and bill allocated PTUs hourly. Azure Reservations discount that meter without securing deployment capacity; throughput and minimum size vary by model.
Amazon Bedrock Provisioned Throughput user guide
Model Units specify input and output throughput across requests, with hourly billing and model-dependent commitments. The documentation directs buyers to an account manager for unit specifications.
Amazon Bedrock, supported Regions and models for Provisioned Throughput
Supported-model list includes open-weight Meta Llama 3.1 70B and 8B Instruct and Llama 3.2 1B, 3B, 11B and 90B Instruct, each shown with us-west-2 single-region support.
Vertex AI Provisioned Throughput overview
Describes Provisioned Throughput as a fixed-cost, fixed-term subscription in several term lengths, bought at a fixed monthly or weekly price with control of overages.
Vertex AI, calculate Provisioned Throughput requirements
GSUs measure prompt and response processing using model-specific burndown rules. Cached input has its own consumption weighting, and unused throughput does not carry over.
Vertex AI, models supported by Provisioned Throughput
The September snapshot lists supported models and model-specific throughput and burndown rates. Equal GSU counts do not represent equal output throughput across models.
Vertex AI, purchase Provisioned Throughput
Orders cannot be cancelled mid-term; monthly subscriptions can renew automatically, weekly terms cannot; overages are billed as pay-as-you-go and controllable per request; orders are processed against available capacity and can take from a few minutes to a few weeks.
Google Cloud Agent Platform pricing, Provisioned Throughput section
The September snapshot lists global-endpoint GSU-hour rates varying by commitment term. Model-specific consumption rules and dated promotional credits must be checked before purchase.
OpenAI Scale Tier for API customers
The September 16 snapshot defines model-snapshot token bundles with a 30-day minimum. Token weighting differs by bundle; the published TPM figure is not universally an output-token allowance.
OpenAI Help Center, Scale Tier for existing enterprise customers
Confirms the token-unit definition and 30‑day minimum; units are added and removed in the API Platform account after an order form is signed, and renew daily.
Anthropic API service tiers documentation
The September snapshot identifies Priority Tier capacity commitments as no longer purchasable. Its model-specific input/output commitments and standard-tier fallback are historical context.
The September 29 snapshot lists H100 and B200 GPU-hour prices and model-dependent PTU token types. PTU model rows were unreadable, so no PTU price is quoted.
The September snapshot bills on-demand deployments per GPU second without startup charges. Its dated hardware price table publishes no token-throughput entitlement.
The September snapshot lists per-minute dedicated hardware rates and no idle charge under the documented rules. Workload capacity depends on the configured deployment and measured performance.
Standard Thinking Product Plan
Listed model output allocations, monthly unit pricing, and extra-capacity billing support the planning examples. Prices, allocation conditions, and service commitments are confirmed before paid use.
Continue exploring
34%utilization where API cost equals one unit
- Model
- GLM‑5.3
- Unit tokens/s
- 135
- API total
- $3,499.20
Calculation example, not a measurement
How to buy inference for an AI product
A buying guide to inference billing models, with a worked break-even example and the service terms to check before committing.
$2,700a month for the example product
- Pay per token
- $4,500
- Base
- 2 GLM‑5.3 units
- Auto scaling
- 60 blocks
Calculation example, not a measurement
Estimate Product Plan output units and cost
Use traffic, answer length, and cached input to estimate monthly output capacity and compare plan cost with token charges.
57.40public index, 8B SH
- 3B SH
- 51.53
- Retrieval & Classification
- 66.4
- Tools & Automation
- 84.0
Decision Index 0.3 public index, self-scored
Standard One SH scores every option in one pass
Open-weight decision models that return a probability for every option in one forward pass: 51.5 and 57.4 on the Decision Index 0.3 public index, self-scored with the official kit, with the detail and the limits.