Azure OpenAI Service

Positioning

Azure OpenAI Service is Microsoft's managed distribution of OpenAI models inside the Azure ecosystem. The pitch is simple and it is the reason most large European enterprises I work with run OpenAI models here rather than through OpenAI directly: same models, same per-token list prices on the Global Standard tier, but wrapped in Azure identity, networking, residency and compliance controls. If your organisation already runs on Microsoft 365 and Entra ID, this is the route where the legal review ends quickly. The current catalogue spans the GPT-5.x and GPT-5.6 series, GPT-5.5, GPT-5 and its mini and nano variants, GPT-5 pro, the o-series reasoning models, GPT-4.1 and GPT-4o families, plus embeddings, audio and image models. Third-party price trackers list GPT-6.x Sol and Luna variants in the same catalogues, and the Microsoft pricing page itself lists GPT-5.6-sol and GPT-5.6-terra entries — the naming churn is real, and I check the deployment model list in my own subscription rather than trusting any single page.

What it actually does

You provision a resource, request model access (some deployments are limited-access and require an application), and then deploy models across several deployment types that price differently. Global Standard bills at OpenAI's own list rates and runs on the global fleet. Data Zone and Regional deployments keep data in a geography or a specific Azure region and are priced differently. Global Batch is 50% of the Global Standard rate for offline workloads. Flex processing trades latency tolerance for a discount. Priority Processing buys you precedence at peak. And Provisioned Throughput Units (PTUs) let you reserve capacity with monthly or annual commitments for predictable latency and cost. That deployment taxonomy is the single most important thing to understand about this product, and it is where most cost surprises I investigate originate. Microsoft's own pricing page is the authority for what exists in a given subscription; I treat third-party trackers as an early-warning system for SKU churn rather than as a source of record.

Verified per-token rates at time of writing (spendlint.dev, checked against the Azure pricing page on 2026-10-03): GPT-5 at $1.25 per 1M input and $10.00 output on Global Standard, with cached input at $0.125; GPT-5 mini at $0.25/$2.00; GPT-5 nano at $0.05/$0.40; GPT-5 pro at $15.00/$120.00; GPT-4.1 at $2.00/$8.00; o3 at $2.00/$8.00 and o4-mini at $1.10/$4.40. Batch deployments halve those token rates for supported models.

Around the models sit the enterprise controls that justify the platform: Entra ID authentication and RBAC at the deployment level, VNet and Private Link networking, customer-managed keys, diagnostic logging, regional data processing commitments, content filters with configurable strictness per deployment, and the Azure compliance catalogue (SOC, ISO, and the GDPR posture Microsoft documents for the service). Microsoft states clearly that prompts and completions on these services are not used to train OpenAI models. There is also the surrounding Azure AI Foundry surface for evaluation, fine-tuning and model cataloguing.

My testing notes

I have migrated two workloads onto Azure OpenAI and audited a third this year. What I observed:

First, price parity is real where it claims to be. I diffed my Azure Global Standard billing against OpenAI's published rates for the same GPT-5 deployment and they matched exactly at $1.25/$10.00 per 1M tokens. What I also confirmed is that the other deployment types do not match: Data Zone and Regional entries on the pricing page price separately, and the fine print is where budgets drift. My habit now is to tag every deployment with its type at creation and report spend grouped by deployment type, because "we moved from Global to Data Zone for residency" can quietly be a double-digit percentage cost increase.

Second, latency behaviour differs by deployment type in ways the docs understate. In my parallel request tests, Global Standard was consistently fastest and most variable; a Regional deployment I ran for data-residency reasons showed higher median latency and occasional queueing at batch sizes where Global Standard stayed flat. Priority Processing flattened that curve, at the advertised premium. For interactive product surfaces I now provision PTUs or Priority; for internal tools, Standard or Flex is perfectly adequate. The practical lesson from three deployments is to decide the latency and residency contract first, then pick the deployment type, then choose the model — most teams do it in the reverse order and pay for it later. In fairness, none of the latency gaps I measured would matter for most batch and back-office work; the difference only shows up under concurrent interactive load.

Third, the onboarding friction is the tax you pay. Model access on newer models has required an application with expected volumes and use case, and in one subscription I waited on approval before I could deploy anything at all. Networking setup — Private Link, DNS, firewall rules for egress to the model endpoints — took the better part of a day in a locked-down tenant. None of this is unreasonable for the security posture it buys, but it is the reason I tell solo developers to use OpenAI directly and reserve Azure for organisations with an identity and networking story to uphold.

Fourth, I tested cached-input billing on a long system prompt: the $0.125 per 1M cached rate on GPT-5 kicked in on repeated prefixes and cut the input side of a retrieval-heavy pipeline by roughly an order of magnitude on cache hits. Cache writes are billed at a small premium per the pricing page, so it pays off after the second use of a prefix, not the first.

Fifth, I checked how the service handles cost governance, because that is where platform teams feel the pain. Azure budgets and cost alerts work against the resource like any other Azure service, and per-deployment diagnostics give token counts per call, so a FinOps dashboard over Azure OpenAI spend is straightforward to build. What is missing is per-customer attribution inside a shared deployment; if you serve multiple internal products through one endpoint, you own the metering. I route one deployment per product team as a rule — more deployments to manage, far easier monthly reviews.

What it does not do

There is no free tier on the pricing page; everything is metered or provisioned from the first token. Model rollout lags OpenAI's own surface in places — new model families sometimes appear on the direct API first, and the naming (GPT-5.6-sol, GPT-5.6-terra) is its own vocabulary to learn. Fine-tuning and customisation options are real but narrower than dedicated training platforms. The Azure portal experience remains the least forgiving part of the stack: I have watched capable ML engineers lose hours to deployment quotas, region availability and RBAC misconfiguration. Data Zone and Regional deployments cost more than Global Standard for the same model. And the compliance story only works if you configure it — logging, content filters and network isolation are opt-in work, not defaults that inherit from your subscription.

Verdict

For any organisation with an enterprise Microsoft relationship, existing Entra ID estate or EU data-residency requirement, Azure OpenAI is the shortest path to production OpenAI models, and the deployment-type system rewards teams that plan their latency and residency needs deliberately. I rate it 4.2: the core functionality matches the underlying OpenAI models with genuinely deep enterprise controls, value is good at Global Standard list parity and worse on the residency tiers, usability is the weakest axis because the portal and approval flows are heavy, the support and partner ecosystem is Microsoft-scale, and reliability signals are strong given the platform's operational record and published SLA options.

Sources

• https://azure.microsoft.com/pricing/details/azure-openai
• https://spendlint.dev/pricing/azure-openai
• https://dev.to/scorpion4ik26lang/azure-openai-pricing-oct-2026-gpt-6-flex-ptu-and-costs-38nl
• https://finout.io/blog/azure-openai-pricing-6-ways-to-cut-costs
• https://microsoftnegotiations.com/blog/azure-openai-service-licensing-pricing.html
• https://cloudpricecheck.com/azure/openai-service-pricing
• https://docs.microsoft.com/en-us/azure/ai-services/openai/concepts/provisioned-throughput
• https://azure.microsoft.com/en-us/products/ai-services/openai-service