AI cost · Microsoft

Azure OpenAI cost: what you actually pay

Token rates mirror OpenAI's direct API almost exactly, so the real cost lever isn't the sticker price — it's Global vs. Data Zone vs. Regional deployment premiums and Provisioned Throughput (PTU) minimum commitments that bill hourly whether or not you use them.

Public pricing

PlanPriceBillingIncludes
Pay-as-you-go (Global Standard)CustomMirrors OpenAI's direct API rates for Global deployments; Data Zone/Regional deployments price separately.usagePay-per-token access to deployed models
Provisioned Throughput (PTU) - monthly reservation$2,448 / per PTU/moApproximate starting rate per third-party analysis; varies by model/region/deployment type. Minimum purchase is 15-50 PTUs depending on deployment.customDedicated reserved capacity with predictable latency
Provisioned Throughput (PTU) - yearly reservationCustomRoughly 35% cheaper than the monthly reservation rate; exact figure requires the Azure pricing calculator for your model/region.customSame capacity as monthly PTU with a longer-term discount
Batch APICustom24-hour turnaround, asynchronous processing.usage

Token rates (USD per 1M tokens)

ModelInputOutputCache read
GPT-4oGlobal Standard deployment; Data Zone/Regional pricing can differ for the identical model.$2.5$10
GPT-4o miniCheapest mainstream general-purpose model in the family.$0.15$0.6
GPT-4.1Cached input ~75% off list. GPT-5 (Global Standard) runs about $1.25/$10 input/output; GPT-5 Pro up to $15/$120.$2$8$0.5

Azure's Global Standard token prices track OpenAI's direct API essentially 1:1 (GPT-4o $2.50/$10, GPT-4o mini $0.15/$0.60, GPT-4.1 $2/$8, GPT-5 ~$1.25/$10 per MTok). Data Zone and Regional deployments can price the same model differently. Provisioned Throughput Units reserve dedicated capacity billed hourly (or via 1-month/1-year reservation, ~35% cheaper annually) regardless of traffic; minimum purchases of 15-50 PTUs mean an idle reservation can run into the tens of thousands of dollars a month. Cached-input discounts (roughly 50-90%, model-dependent) only help prompts with a genuinely reused prefix.

Approximate — verify before relying on it. Last updated: 2026-07-22. Sourced from Azure OpenAI Service Pricing (official), Microsoft Learn: Provisioned Throughput (PTU) costs and billing.

Where the real Azure OpenAI cost hides

The sticker is the smallest part of the story. For Azure OpenAI, these are the line items that quietly inflate the bill:

  • GPT-4o ($2.50/$10) vs. GPT-4o mini ($0.15/$0.60) differ by ~17x — defaulting new features to the full-size model instead of mini/nano tiers is the most common overspend.
  • Regional and Data Zone deployments can carry different per-token prices than Global Standard for the identical model — teams price-check once at launch and don't notice the delta after a region migration.
  • PTUs bill by the reserved unit-hour for the entire commitment term regardless of whether traffic shows up — minimum purchases (15-50 PTUs) can cost tens of thousands per month sitting mostly idle during ramp-up.
  • Output tokens price 4-8x input on nearly every model, and reasoning models (o-series, GPT-5 Pro) add hidden 'reasoning tokens' billed as output even when never shown to the user.
  • Cached-input discounts require a stable, reused prompt prefix within the cache TTL — apps with per-request-unique system prompts get none of the caching discount they assume is already applied.

What an audit finds on Azure OpenAI

Teams provision GPT-4o (or GPT-5) globally for every workload, skip the mini/nano tiers and prompt caching, and either overpay on-demand by 15-20x for simple tasks or lock in a PTU commitment before traffic volume justifies it.

Azure OpenAI pricing FAQ

Is Azure OpenAI more expensive than OpenAI's own API?

Per-token list prices for Global Standard deployments are effectively the same as OpenAI's direct API; Azure's premium shows up in PTU minimums, Data Zone/Regional surcharges, and the surrounding Azure bill, not the token rate itself.

What's a PTU and when do I need one?

A Provisioned Throughput Unit reserves dedicated model capacity billed hourly regardless of usage. It only pays off at sustained volume (commonly cited around 150M+ tokens/month) — below that, pay-as-you-go is cheaper.

Does prompt caching apply automatically?

Yes for supported models, but only when the prompt has a stable, reused prefix; cache hits run 50-90% cheaper than base input, cache misses bill at full price.

What is Azure OpenAI really costing you?

Spendassay connects your Azure OpenAI spend to actual usage and engineering output — a proof-level AI cost report with the wasted dollars named. Neutral across every AI vendor.

Free · read-only · no card

More: the complete AI cost management guide · every tool, with prices · cost per outcome · compare platforms · how we measure