Azure OpenAI Pricing 2026 – Is PTU Worth the Cost?
Oct 30, 202614 min read
Three ways Azure OpenAI charges
Azure OpenAI pricing and whether PTU pays off
Per token, per hour, per term: Azure OpenAI bills in three ways, whether you call it directly or from a workflow built in a tool like Activepieces.
They are per million tokens on Standard and Batch, per hour on hourly Provisioned Throughput, or per month or year on a reservation of provisioned throughput units (PTUs).
Azure OpenAI pricing is the set of billing models for using OpenAI models on Azure, which charges either per token on pay-as-you-go Standard and Batch deployments or for reserved model capacity in provisioned throughput units (PTUs), billed hourly or by monthly or annual reservation.
Multiply input tokens by the input rate, add output tokens times the output rate, and you have the cost of one run on pay-as-you-go.
Provisioned pricing breaks that link. You pay for capacity whether or not you use it, so the cost per run depends on how full you keep it.
Azure OpenAI pricing check date and limits
As of 30 October 2026, the structure above reflects Microsoft's Azure OpenAI pricing page and its provisioned throughput documentation.
Treat it as a snapshot: Microsoft changes model availability, regional offers and reservation terms, so the page you open will be the authority on the day you buy.
You will not find per-token or per-PTU rates here. They would be stale by the time you read them, and the right figure depends on model, region and deployment type.
If you are running this arithmetic for your own team, see what the same workload costs on Activepieces.
Six criteria for judging an Azure OpenAI plan
Judge any Azure OpenAI plan on price per unit, inclusions, limits, performance guarantees, commitment terms and predictability, because those six decide whether pay-as-you-go or provisioned throughput costs you less.
- Price per unit: what does one token cost me, and does that rate change with the model or the deployment type?
- What is included: which charges sit outside the headline rate, and what do I still pay for if I use the plan lightly?
- Rate and quota limits: how much traffic can I send before requests are refused?
- Latency and throughput guarantees: is response speed promised, or merely typical?
- Commitment and cancellation terms: how long am I bound, and what does leaving cost?
- Cost predictability: can I forecast next month's bill within a margin I could defend to finance?
Price and what it includes
Divide the full monthly charge by the work actually done, not the work you reserved.
Under reserved capacity, unused headroom still appears on the invoice, so the effective price per request can sit well above the listed one.
Azure OpenAI quotas, latency and guarantees
Do not read quota as a promise of speed. It is a ceiling.
Because pay-as-you-go shares capacity with other customers, latency can vary.
Reserved capacity exists to remove that variance, and it's the only one of these criteria where it clearly earns its price.
Azure OpenAI PTU commitment and predictable cost
Commitment buys predictability, and you pay for it in flexibility. A fixed monthly charge is easy to forecast but unforgiving if volume falls.
Commitment buys predictability, and you pay for it in flexibility.
Every current Azure OpenAI plan in one table
Eight kinds of capacity are on sale from Azure OpenAI. Only the pay-as-you-go ones are plans where you can multiply a published rate by your own token counts and get a bill.
Read against Azure's pricing page on 2026-10-30, the table below records the billing basis, inclusions and limits for each plan.
| Plan | Price basis | Inclusions | Usage limits |
|---|---|---|---|
| Standard | Per token, input and output billed separately | Processing in the region you deploy to | Per-model token quota per subscription and region |
| Global Standard | Per token | Requests may be processed in any Azure region | Its own quota pool, separate from Standard |
| Data Zone Standard | Per token | Processing stays within a defined zone, such as the US or the EU | Its own quota pool |
| Batch | Per token, discounted against the online rate | Asynchronous jobs with a target completion window | Separate queued-token quota |
| Provisioned Global | Hourly per PTU; reservations lower the rate | Reserved throughput, processed in any region | Minimum PTU purchase varies by model |
| Provisioned Data Zone | Hourly per PTU; reservations lower the rate | Reserved throughput, processing kept in a zone | Minimum PTU purchase varies by model |
| Provisioned Regional | Hourly per PTU; reservations lower the rate | Reserved throughput in one region | Minimum PTU purchase varies by model |
| Fine-tune hosting | Hourly per deployed custom model, plus training | A dedicated deployment of your tuned model | Not published |
Prices and plan limits checked against docs.claude.com and claude.com and openai.com and gemini.google on October 30, 2026.
Only the first four rows scale to zero when traffic stops, which means you're still paying for the remaining rows when there's no traffic.
Standard, Global Standard and Data Zone Standard bill identically per token. They differ only in where your data may be processed, so the choice is a compliance question first and a price question second.
For work that can wait, Batch is the cheapest per-token route. Fine-tune hosting is the easiest charge to forget: it bills hourly while the model sits deployed, even if nobody calls it.
What each plan costs at realistic monthly usage
In the GPT-4.1 workloads priced by these sources, Standard and Batch cost less than PTU. PTU bills for reserved capacity whether or not you fill it.
Mid volume: where PTU starts to compete
Take a 360-million-token month. In the worked example published in (source), which uses Azure's list pricing, Batch comes to $1,200 and Standard to $2,400, against $3,900 for reserved PTU, making Batch the most cost-effective option for this workload.
Even at this volume, the reservation costs about 1.6 times Standard, so you pay a premium for guaranteed latency you may not need.
Price any job that can wait on Batch first, which is half of Standard, and PTU is not yet in the running.
High steady volume: PTU and reservation math
Move to one million requests a month. Cld2day lists Batch at $2,700 and Standard at $5,400.
A one-year PTU commitment costs $14,365, about 2.7 times Standard, so a million requests is still not enough volume. Month-to-month PTU costs $16,900, which means flexibility adds $2,535 over the annual term.
Work from the 360-million-token figures and the break-even follows. Reserved capacity beats Standard only when Standard spend would exceed the $3,900 reservation.
That's about 1.6 times the volume, or roughly 585 million tokens a month, and the traffic must arrive steadily enough to keep the capacity busy.
The same $2,400 of Standard spend uses only about 62% of that break-even point, and anything spikier leaves paid-for capacity idle, which means provisioning costs outpace actual usage.
So PTU pays off when volume is high, constant, and latency-sensitive all at once. Missing any one of the three, the per-token plans cost less.
Worth checking against a plan that does not meter every step: one credit covers a whole run on Activepieces.
Hidden costs and limits Azure OpenAI buyers hit
Once you count quota ceilings, idle reserved capacity, and the services that sit around the model, the headline per-token rate is the smallest part of the bill.
Azure OpenAI quota limits and throttling
Go over the ceiling and you get a 429, because pay-as-you-go deployments run under tokens-per-minute quotas set per region and model. That means your cost per successful run includes the retries, the backoff delay, and the engineering time spent writing both.
Paying for idle PTU capacity
Whether or not traffic arrives, provisioned throughput bills for the reservation.
Minimum deployment sizes make this worse, because a workload that needs a fraction of the smallest unit still buys the whole unit.
Do the arithmetic on average utilization, not peak. If you can't keep the reservation busy most of the day, the per-token plan is still cheaper.
Few production deployments use the model alone. Each of these arrives as its own line on the invoice:
- Content safety filtering and moderation calls, which add cost and latency to every request they screen.
- A search index for retrieval, billed for provisioned capacity whether or not anyone queries it.
- Private endpoints and network egress, needed once security review demands isolation.
- Log storage and monitoring, which grow with every prompt and completion you retain.
- Hosting for fine-tuned models, charged for the time the deployment exists, not for the requests it serves.
Tracking Azure OpenAI spend per workflow
Azure reports cost by resource, not by the business process that caused it. Two workflows sharing one deployment look like a single line, so you can't say which one justifies a move to PTU.
The admin console of a workflow automation tool, shown below, lists Usage and Billing & subscription in the same sidebar as Connections and Integrations.
Spend visibility works best where the workflow is defined, because that is where a run can be tied to a cost. Tag each workflow's calls, log token counts per run, and reconcile them against the invoice monthly.
Activepieces can add that per-workflow logging and run history and log streaming around your Azure OpenAI calls, and it leaves the model choice and the price with you.
Azure OpenAI against OpenAI API and Amazon Bedrock
What reaches your invoice is how each reservation is bought and what it commits you to, since all three bill per token on demand and sell reserved capacity. The table sets them side by side on six criteria.
Rates change by model and region, so the table leaves out a price per million tokens on purpose. Take your logged token counts to each pricing page for the model you actually run.
| Per-token price | Discounts | Reserved capacity | Rate limits | Data residency | Compliance | |
|---|---|---|---|---|---|---|
| Azure OpenAI | Pay-as-you-go by deployment type | Batch deployments | Provisioned throughput units on term reservations | Quotas per region and subscription | Regional, data zone and global deployment types | Azure certifications |
| OpenAI API | Pay-as-you-go by model | Batch API | Priority processing; larger commitments by contract | Usage tiers that rise with spend | Regional options for eligible customers | OpenAI's own attestations |
| Amazon Bedrock | On-demand, set per model and provider | Batch inference | Provisioned throughput on commitment terms | Quotas per region and model | Region choice; cross-region inference can route elsewhere | AWS compliance programs |
Because Bedrock hosts models from several providers, "same model class" is a judgment call and the comparison is only as good as your match.
Batch discounts exist on all three, but they apply to work that can wait, so they only help workflows that tolerate delay.
OpenAI's route to larger commitments runs through a sales conversation, which means you can't model it from a public page.
Limits, data residency and compliance
Azure and Bedrock quotas attach to a region, so a spike in one region doesn't borrow headroom from another. OpenAI's tiers rise with spend, which means a new account starts constrained.
On residency, check the deployment type: Azure's global option and Bedrock's cross-region inference can both process requests outside the region you picked. Compliance follows the platform, so ask which certifications your auditor already accepts.
What Activepieces does about this
The decision in this article rests on one number most teams don't have: steady token volume per workflow. Activepieces helps you get it. Each Azure OpenAI call runs as a step inside a flow, and every run is logged with its history.
You can tag a flow, record token counts per run, and set the total against your invoice. That tells you whether a flow's volume is anywhere near the roughly 585 million tokens a month where reserved capacity beats Standard, or nowhere close.
It also stays out of the price. Activepieces calls the deployment you pick with your own Azure key, so tokens bill at your Azure rate and land on your Azure invoice.
With your own AI keys, each AI step counts as 1 credit from the $20 Plus plan up, so per-step charges only apply once you've committed to a paid plan.
The Ultimate plan breaks credit usage down by project, which shows which team is driving spend. The core is open source under the MIT licence, so you can read how a step is billed before you build on it.
Moving the work that can wait to Batch, and keeping the rest on Standard, is a change to the deployment a flow points at, not a rebuild.
In the worked example above, that is the difference between $1,200 and $2,400 for the same 360 million tokens, against $3,900 reserved, so the pricing model you select will significantly impact your total costs for workloads at this scale.
Run the flows on pay-as-you-go, keep the logs, and revisit PTU only when those logs show volume that is high, constant and latency-sensitive all at once.
Activepieces is the better fit for teams that want to decide their own AI and Azure OpenAI strategy, judged on cost transparency and control over which model deployment each workflow uses.
It calls the deployment you choose with your own Azure key, so tokens bill at your Azure rate rather than at a resold markup, and each AI step counts as 1 credit from the $20 Plus plan up.
Its per-run logs let you measure steady token volume per flow before committing to reserved capacity, and moving a flow between Batch and Standard is a change of deployment rather than a rebuild.
Frequently asked questions
Most of what is left is billing mechanics, and the answers are less dramatic than a sales deck makes them.
Is Azure OpenAI more expensive than OpenAI directly? Not on the per-token rate for the same model, which generally tracks the direct price.
The difference shows up around the tokens: private networking, logging and monitoring, and any regional or data-zone deployment you choose for compliance all add lines to the invoice. Price the whole stack, not the model line.
Can I switch from pay-as-you-go to PTU without downtime? Yes, because a provisioned deployment is a separate deployment alongside your existing one. Nothing forces a cutover.
Stand up the PTU deployment, send a slice of traffic to it, and keep the pay-as-you-go deployment as overflow. The risk is capacity, not downtime: confirm the region can provision what you need before you commit to a reservation.
Do Azure credits or an enterprise agreement cover Azure OpenAI? Often, since it bills as an Azure service against your subscription, so a negotiated commitment can absorb it.
Whether a specific credit or discount applies depends on the terms of your agreement, so get the answer in writing before you plan around it. A pricing answer that only comes on a call is not one you can budget against.
How do I set a spending cap on Azure OpenAI?
There's no hard dollar cap on the resource itself. The practical controls are layered:
- A budget in Cost Management with alerts, wired to an action group that can notify or run automation.
- A tokens-per-minute quota on each deployment, which limits how fast you can spend.
- Token limits at a gateway in front of the service, if you need per-team or per-key ceilings.
The quota is the only one that stops spend as it happens, so set it deliberately.
Here's the whole decision as five steps you can finish on Monday, each one feeding the next:
- Pull 30 days of token usage by deployment, so you decide on a full month of traffic and not a lucky week.
- Compare peak tokens per minute against your quota. A peak that sits far below quota means you're not capacity-constrained.
- Price the same usage on Standard and on Batch. Any work that can wait for results moves to the cheaper rate immediately.
- Test PTU break-even by dividing the reservation cost by your steady hourly usage at Standard rates. If your real utilization never reaches that line, reserved capacity costs more.
- Decide. If the math doesn't clearly favor PTU, stay on pay-as-you-go and rerun the test next quarter.
Related reading
References
Running the numbers
See what the same workload costs here.
Free forever plan, and every paid plan self-hosts at no extra cost.
See pricing Talk to sales
