Azure's sizing docs give you an exact formula for converting requests-per-minute into Provisioned Throughput Units. Azure's own pricing page doesn't give you the number to multiply that formula against: every PTU row shows a literal "$-" with a note that prices are estimates only, and tells you to sign in to the pricing calculator or talk to an Azure sales specialist for your real rate. That's the actual shape of Azure OpenAI PTU pricing in 2026, exact sizing math paired with a rate you can't see until you're inside an active account.
TL;DR: Azure OpenAI PTU Pricing vs a Fixed-Price Inference Plan
Azure OpenAI doesn't publish a flat per-PTU rate; it publishes the billing structure and sizing math, leaving the price to a sales call.
| Factor | Azure OpenAI PTU | Spheron fixed-price plan |
|---|---|---|
| Billing | Hourly per PTU, billed whether idle or not | Prepaid tokens/month, price fixed for the term |
| Minimum | 15 PTUs Global/Data Zone, 50 Regional | 5B tokens/month on Starter, no PTU math |
| Rate | Shown as "$-"; quote-only, no public number | Quoted per plan, not published |
| Capacity signal | PTU count needs a TPM-per-PTU conversion | Concurrency stated directly: 8 or 32 at once |
| Verdict | Wins on sustained traffic against a quoted rate | Wins when you want concurrency stated upfront |
See current GPU pricing for live rates.
How Azure OpenAI PTU Pricing and Minimums Actually Work
Provisioned Throughput Units are a capacity reservation, not a per-token meter. All provisioned deployment types are billed at an hourly rate ($/PTU/hr) based on the number of PTUs deployed, regardless of the number of tokens consumed. That's the opposite of the pay-as-you-go tier most teams start on, where idle time costs nothing and a traffic spike just gets throttled instead of billed extra.
The catch is the minimum deployment size. Global and Data Zone Provisioned deployments require a minimum of 15 PTUs for most current models, including gpt-5.2, gpt-5.4, gpt-6-astra, and gpt-6.1-sol, scaling upward in increments of 5 PTUs. You can't provision 3 PTUs to test a model at small scale; you buy 15 or you stay on pay-as-you-go. Regional Provisioned deployments set the floor even higher: a minimum of 50 PTUs for most models, though several mini and nano variants, gpt-5.4-mini, gpt-4.1-mini, gpt-4.1-nano, and o4-mini, only require 25.
None of that tells you what 15 PTUs is actually worth in requests per minute, because capacity-per-PTU is model-specific. gpt-5.2 supports 3,400 input tokens per minute per PTU, with output tokens counted at 8x the cost of input tokens toward that same capacity on Global and Data Zone Provisioned deployments. A different model ships a different TPM-per-PTU number and a different output multiplier, so the same 15-PTU reservation buys meaningfully different throughput depending on which model you deploy it against.
Azure OpenAI PTU Pricing: What a 15-PTU Deployment Costs Hourly, Monthly, and Reserved
Azure structures PTU billing around three windows, and the 15-PTU Global/Data Zone minimum sits under all three the same way:
| Billing window | Commitment | How the rate is set |
|---|---|---|
| Hourly billing | None, cancel anytime | $/PTU/hr; no rate published publicly |
| Monthly reservation | 1 month | Discounted $/PTU/hr vs hourly; no rate published publicly |
| Yearly reservation | 1 year | Deepest discount vs hourly; no rate published publicly |
Azure's own pricing page lists "Per PTU Hourly pricing," "Per PTU Monthly Reservation Pricing," and "Per PTU Yearly Reservation Pricing" as column headers for every supported model and region, with the minimum PTU count shown for each row. Every dollar figure in those columns renders as a dash, and the page states plainly that "Prices are estimates only and are not intended as actual price quotes," recommending you "Sign in to the Azure pricing calculator to see pricing based on your current program/offer with Microsoft" or "Contact an Azure sales specialist for more information on pricing or to request a price quote."
The reservation mechanic itself is public even though the number isn't. Azure Reservations are a financial discount applied to the PTU billing meter, and in exchange for a 1-month or 1-year commitment, you receive a discounted effective $/PTU/hr rate. What that means in practice: the structure of the discount is documented in detail, hourly costs more per unit than monthly, monthly costs more per unit than yearly, but the actual $/PTU figure your 15-PTU deployment bills at depends on your account's program and offer, not on anything posted at a URL you can read without signing in.
Pricing for Azure OpenAI Provisioned Throughput varies by program, offer, and region, and isn't published as a flat rate on Azure's own pricing page as checked on 31 Oct 2026; use the Azure pricing calculator or an Azure sales specialist for your actual $/PTU rate before committing to a reservation. This post quotes no Spheron GPU rate directly, since the comparison here is against a token-based inference plan, not an hourly rental rate.
Matching PTU Capacity to Counted Tokens on a Fixed Plan
A fixed-price inference plan skips both of those steps by stating the guarantee in the unit a buyer actually cares about, at a price fixed up front. Spheron's managed inference plans are prepaid, fixed-price allotments of counted tokens on reserved GPUs, never spot, with the price fixed for the term rather than tracking the GPU market. Dedicated starts at 50B tokens a month on single-tenant nodes pinned to the EU or US, for teams running custom or fine-tuned open models. Token counting weights output at 4x a fresh input token and cached input at 0.25x (0.03x on DeepSeek V4.1 Flash), which rewards the same kind of cached, repeat-context agent traffic that Azure's own sizing formula also treats as cheaper to serve.
The honest gap: this comparison only holds if the model fits. Spheron's plans currently run GLM-5.3, GLM-5.3-Flash, and DeepSeek V4.1 Flash, not GPT-6 Astra, GPT-6 Sol, or any other OpenAI model family. A team already committed to an OpenAI model has no path off PTU through this plan. And like Azure's own PTU rate, plan pricing is quoted rather than posted on a page, so there's still no public dollar figure to set side by side in a table; both paths end at a request form before you see a number.
Which Model Wins at Your Traffic Volume
The honest answer is that PTU and a fixed-price plan are solving the same underlying problem, predictable capacity instead of per-token meter risk, but exposing the sizing decision at different layers, and neither hands you a public number without a conversation.
PTU wins when your traffic is large and steady enough to keep the reservation full and you have an active Azure relationship to quote against. If you're already well past the 15-PTU minimum and your request volume holds roughly flat hour to hour, a yearly reservation brings the effective cost per PTU down from the hourly rate, and you're paying for Azure's full GPT model family directly rather than a narrower open-model lineup.
A fixed-price plan wins when you'd rather commit to a concurrency and token ceiling you can read off a spec sheet than run a TPM-per-PTU formula per model before every sizing decision, and when your model choice isn't locked to OpenAI's lineup. Spheron's Growth plan guaranteeing 32 concurrent requests is a number you can hold against your own expected concurrent users directly, without first translating it through tokens-per-minute math.
If your traffic is small, bursty, or still finding its shape, neither reserved option is where to start. Azure's own pay-as-you-go tier and Spheron's Starter plan both exist for exactly that stage, where the cost of guessing wrong on a 15-PTU or 5B-token commitment outweighs any discount a reservation offers. For a broader look at where serverless, on-demand, and reserved billing each make sense before you lock into either PTU or a fixed plan, see Spheron's breakdown of the three GPU billing models. AWS runs the same structural trade under its own name: Bedrock's Provisioned Throughput bills a fixed hourly Model Unit rate regardless of use, making it the closer comparison for teams evaluating more than one hyperscaler's reserved-capacity tier. And for the crossover math behind when per-token billing stops being the cheaper option entirely, Fireworks AI's pricing against renting dedicated GPU capacity walks through the same question from the self-hosting side.
If your sizing math keeps landing on a concurrency number rather than a raw PTU count you still have to price out, a fixed-price plan that states that number directly might save you the recalculation every time traffic shifts.
Frequently Asked Questions
Azure doesn't publish a flat number. Its own pricing page shows a placeholder dash for every Provisioned Throughput tier and states prices are estimates only, directing you to the Azure pricing calculator or an Azure sales specialist for your program's actual rate. What is public: PTUs bill hourly per unit regardless of token volume, with monthly and yearly reservations available that discount that hourly rate for a term commitment, and a 15-PTU minimum on most current models for Global and Data Zone deployments.
15 PTUs for Global and Data Zone Provisioned deployments on most current models (gpt-5.2, gpt-5.4, gpt-6-astra, gpt-6-sol, and others), scaling in increments of 5. Regional Provisioned deployments need 50 PTUs for most models, though several mini and nano variants (gpt-5.4-mini, gpt-4.1-mini, gpt-4.1-nano, o4-mini) only need 25.
It depends entirely on sustained utilization and on the rate Azure quotes your account, which isn't published. PTU bills per deployed unit per hour whether or not you send a single request, so idle capacity is pure loss. Pay-as-you-go bills per token with no minimum and no idle cost. PTU only wins once sustained traffic consistently fills the capacity you reserved, which means running the sizing math against your own quoted rate before committing, not estimating it from a public price list that doesn't exist.
Run your peak requests-per-minute, average prompt tokens, and average completion tokens through Microsoft's PTU sizing formula, which varies by model because each model supports a different tokens-per-minute-per-PTU rate and a different output-to-input cost ratio. Microsoft's own worked example for gpt-5.2 at 1,000 peak RPM needed 110 PTUs rounded up at 0% prompt cache, or 80 PTUs at 50% cache, which is a useful sanity check for the order of magnitude your own traffic will land in.
