The short answer: per-token prices are identical on both platforms, but Azure OpenAI's total cost of ownership runs 15–40% higher. The direct API wins on pure price; Azure wins on compliance, data residency, and — for sustained workloads — Provisioned Throughput Unit economics.
Token Pricing Is Identical
Start with the fact that surprises most buyers: on a per-token basis, Azure OpenAI and the OpenAI direct API charge exactly the same rates for the same models. GPT-4o is $2.50 per million input tokens and $10.00 per million output tokens on both platforms. GPT-5-class models land at $1.25 input / $10.00 output on either side. Microsoft does not mark up the model — it sells the identical OpenAI model under its own commercial wrapper.
So if the sticker price is the same, why does the Azure bill come in higher? Because tokens are only one line on the invoice. The Azure OpenAI vs OpenAI direct cost comparison is really a comparison of everything that sits around the model. That distinction is also the heart of our broader analysis of AI token pricing and how to negotiate it.
Where the 15–40% Azure Gap Comes From
Azure layers six auxiliary charges on top of identical token rates. Each is individually defensible; together they push the total cost of ownership 15–40% above the direct API for the same workload.
| Cost layer | Azure OpenAI | OpenAI Direct |
|---|---|---|
| GPT-4o tokens (in / out) | $2.50 / $10.00 per 1M | $2.50 / $10.00 per 1M |
| Production support plan | $100–$1,000 / month | Included / usage-based |
| Data egress | $0.087/GB after 100GB free | Not separately billed |
| Fine-tuned model hosting | $1,200–$2,160 / month per model | Hosting bundled differently |
| Private Link / network isolation | Billed add-on | Not offered at this level |
| Log Analytics ingestion | ~$2.30/GB | Not separately billed |
The fine-tuning line is the one that catches teams out: a fine-tuned model on Azure incurs roughly $1,200–$2,160 per month in hosting regardless of how often it is called. A team that fine-tunes three models for occasional use can be paying $5,000+ a month before a single production token flows. We cover this trap in detail in our guide to AI fine-tuning costs and contract terms.
When Azure PTUs Flip the Maths
Pay-as-you-go is not the only Azure deployment model, and for high-volume workloads it is the wrong one. Provisioned Throughput Units (PTUs) let you reserve dedicated capacity at an hourly $/PTU/hr rate rather than paying per token. For sustained, predictable traffic the economics invert.
PTU pricing starts at roughly $2,448 per month and breaks even against pay-as-you-go at around 150–200 million tokens per month for GPT-4o. Below that, pay-as-you-go always wins. Above roughly 50,000 tokens per minute average throughput — or 80%+ average utilisation — annual PTU reservations deliver 40–60% savings versus pay-as-you-go, and an annual commitment saves a further ~35% over the monthly reservation. This is the one scenario where Azure can be materially cheaper than the direct API on a like-for-like workload. Model it carefully: PTUs you don't keep busy are pure waste, which is exactly the usage-based pricing risk we warn buyers to cap.
The MACC Drawdown Advantage
There is a second reason large enterprises route AI spend through Azure even at the premium: Azure OpenAI consumption draws down a Microsoft Azure Consumption Commitment (MACC). If you have already committed to, say, $10M of Azure spend over the EA term, Azure OpenAI usage is funded from that pre-committed budget rather than appearing as net-new operating expense. For a CIO managing to a fixed commitment, that can make Azure the lower marginal-cost option even after the auxiliary fees — and growing AI consumption data strengthens the case for a higher MACC discount tier at renewal ($10M MACC commonly lands 18–22% versus 12–15% at $5M). This is where the AI decision intersects the Microsoft commercial relationship covered across our Microsoft vendor intelligence hub.
How to Negotiate Each Platform
Both platforms are negotiable above a threshold — and below it, published pricing is largely firm. Knowing where that line sits is half the battle.
On Azure, the EA committed price for PTUs is the lever. Enterprises with $5M+ annual MACC commitments regularly achieve 25–35% below list on PTU reservations, and PTU reservations themselves roll into MACC drawdown, compounding the commercial value. Bring current consumption data to the table — it justifies a higher commitment tier and a deeper discount.
On the OpenAI direct side, the inflection point is roughly $500K of annual committed spend. Above it, OpenAI's enterprise team has real discount authority and buyers secure 28–38% below API list, plus commercial protections — data isolation, IP ownership confirmation, SLA upgrades — that standard API terms omit. Below $500K, expect 15–20% for moving monthly to annual and little more. We map the full structure in OpenAI enterprise pricing and negotiation and in the planned FAQ on how OpenAI enterprise is priced. The strongest lever on either platform is credible competitive pressure — having Anthropic, Google Gemini, and the alternate Microsoft/OpenAI path live in the same evaluation, a tactic we detail in comparing GPT, Claude and Gemini enterprise licensing. Whether AI token prices are negotiable at all comes down to which side of these thresholds you sit on.
The Decision Framework
For a regulated enterprise — healthcare, financial services, government — Azure OpenAI is rarely optional. SOC 2, HIPAA, FedRAMP, Private Link isolation, and Regional or Data Zone deployment scopes that pin data to a defined geography are controls the direct API does not match. The 15–40% premium is the price of compliance, not a luxury. For teams without those requirements, and especially for lower-volume or experimental workloads, the direct API is cheaper end-to-end and faster to start. The crossover comes at scale: once a workload is large enough to justify PTUs, or once it can be funded from an existing MACC, the Azure premium narrows or disappears. For a structured way to run this analysis across vendors, see the complete guide to AI procurement, download our AI Procurement Checklist, or request a confidential briefing to benchmark your own AI spend against the market.