LLM Provisioned Throughput and Service Tier Planner
Find how many Azure PTUs or Vertex GSUs you need, when they pay off, and what Flex or Batch saves.
Model
The busiest hours
The rest of the month
Pay-as-you-go tiers
The rest, 60%, on Standard.
- Best reserved plan
- $8,563.60
- Your tier mix
- $8,929.62
Each term, a month
| Term | Cheapest plan | Cover the peak |
|---|---|---|
| Hourly, no reservation$730.00 a PTU, break-even never | $9,399.60no PTUs | $51,100.0070 PTUs51% used |
| 1-month reservation$260.00 a PTU, break-even 98.9% | $9,343.6020 PTUs100% used | $18,200.0070 PTUs51% used |
| 1-year reservation$221.00 a PTU, break-even 84.1% | $8,563.6020 PTUs100% used | $15,470.0070 PTUs51% used |
A fully used PTU serves 131M weighted tokens a month, worth $262.80 on pay as you go at this mix. That is what each term’s price is weighed against.
What a token served from reserved capacity costs
Per million weighted tokens, the unit an uncached input token is counted in. The FinOps Foundation’s cost and usage tracker paper estimates it as the full-use rate times (2 minus utilization). A unit costs the same however few tokens it serves, so the bill works out to the full-use rate divided by utilization. The two agree at 100% and part below it. On pay as you go they cost $2.00.
| Plan | Foundation formula | What it costs |
|---|---|---|
| Cover the peak, 1-year reservation51.1% used | $2.50 | $3.29 |
| Cheapest plan, 1-year reservation100.0% used | $1.68 | $1.68 |
Pay-as-you-go tiers, a month
| Tier | Share | A month |
|---|---|---|
| Standard, 1xThe list price everything else is a multiple of. | 60% | $5,639.76 |
| Priority processing, 2xServed ahead of Standard traffic, for lower and steadier latency. | 10% | $1,879.92 |
| Flex, 0.5xSlower, and can be turned away when capacity is short. Not for anyone waiting on the answer. | 30% | $1,409.94 |
| Batch, not offeredAzure lists no Batch price for this model. | 0% | $0.00 |
Rates used and their sourcesShow
- Azure provisioned throughput prices, source, read 17 Oct 2026
- Azure PTU throughput per model, source, read 17 Oct 2026
- Azure OpenAI pay as you go and tiers, source, read 17 Oct 2026
- OpenAI API service tiers, source, read 17 Oct 2026
- Vertex AI provisioned throughput and tiers, source, read 17 Oct 2026
- Vertex AI throughput per GSU, source, read 17 Oct 2026
- Amazon Bedrock service tiers, source, read 17 Oct 2026
- Anthropic API service tiers, source, read 17 Oct 2026
What this model assumes
Show
- Prices are for Global deployments: Azure Global Standard and Global provisioned, the Vertex global endpoint, Bedrock global cross-Region. Data Zone, regional and US-only processing cost more.
- A month is 730 hours, as the vendors price it. The month is two levels of load: the busy hours, and a lower load the rest of the time.
- Tokens past the reserved units are priced at pay-as-you-go Standard, with the same mix of input, cached input and output as the rest. Vertex bills overage that way; Azure does when spillover to a Standard deployment is set up, and returns a 429 when it is not.
- Capacity is counted in the weighted tokens each provider publishes per model: an output token uses 5 or 6 times the capacity of an input token, and cached input little or none.
- Cache writes are counted as ordinary input, and long-context prices (prompts over 200,000 to 272,000 tokens) are left out.
- Tier multipliers apply to every token type. A tier the model does not offer is priced as Standard.
- Promotional credits are applied where the provider states them. Negotiated discounts, Azure savings plans and Google Flexible Savings Plans are not.
Prices verified 17 Oct 2026. Public list prices in USD for Global deployments, before any discount you have negotiated.
Signing a PTU reservation or a GSU term?
Commitment Management: Each month we size your Savings Plans, Reserved Instances and committed use discounts; you approve and buy. From $500 a month, or 12%. Nothing paid upfront.
About this planner
Azure OpenAI and Vertex AI sell LLM capacity by the unit, as PTUs and GSUs, for an hour, a week, a month or a year. Every platform also sells pay-as-you-go tiers above and below Standard: faster at a premium, or slower at half price. Both are ways to change the same bill, and they suit different traffic.
Describe the load at your busiest hours and how the rest of the month compares. The planner finds the number of units that costs least, with the peaks spilling to pay as you go, shows the utilization a unit needs on each term to pay for itself, and prices the share of traffic you could move to another tier.
Questions
When is provisioned throughput cheaper than pay as you go?
Only when the capacity is busy most of the month. For current models a fully used unit costs about what the same tokens cost on pay as you go, so an Azure PTU on a 1-year reservation pays for itself at about 84% utilization and on a 1-month reservation at about 99%, and a Vertex GSU on a 1-year term at about 72% to 75%. Below that, reserved capacity buys guaranteed throughput and steadier latency, not savings.
How many PTUs or GSUs do I need?
Enough for the load you want guaranteed, counted in weighted tokens: an output token uses 5 or 6 times the capacity of an input token on current models, and cached input little or none. Sized for the busiest hour, units sit idle the rest of the day; the cheapest plan usually covers the steady load and lets the peaks spill to pay as you go.
How does this differ from the FinOps Foundation formula?
The Foundation’s paper on building a generative AI cost tracker estimates the cost of a token served from reserved capacity as the full-use rate times (2 minus utilization). That allocates cost fairly between teams sharing the capacity, but it understates the bill below full use: at 50% it gives 1.5 times the full-use rate, while units that cost the same however busy they are work out to 2 times. The planner shows both.
Which pay-as-you-go tiers does each platform offer?
OpenAI: Fast mode (formerly Priority) at 2 times Standard, Flex and Batch at half. Azure OpenAI: Priority at 2 times and Flex at half on the GPT-6 and GPT-5.6 models. Vertex AI: Priority PayGo at 1.8 times and Flex at half for Gemini, Batch at half for every model. Bedrock: Priority at 1.75 times, Flex and Batch at half where the model supports them; the Claude 5.5 models are Standard only. Anthropic API: Batch at half.
What about Bedrock Reserved, OpenAI Scale Tier or Anthropic Priority Tier?
None can be priced from public figures for current models. Bedrock’s Reserved tier covers Claude 4.x and earlier, and its 1-month price works out to Standard at full use. OpenAI sells Scale Tier through sales with no published price. Anthropic no longer sells Priority Tier commitments.
Related tools
LLM Token Cost Calculator
Compare monthly AI spend on Claude, GPT, Gemini and more before the invoice arrives.
OpenAI Usage Export Analyzer
Drop in your OpenAI or Claude usage export and see where the AI spend goes.
OpenAI Agent Cost Calculator
Agents make many model calls per task. Price the whole loop, not one request.
OpenMore tools for: AI and model usage
Reserved capacity only saves money while it is busy.
Finitizer Tokenomics tracks AI spend and usage by model, team and hour, so idle PTUs and traffic that could run on Batch show up before the next reservation renews.
See how Finitizer works