Two sentences about the same price, and only one of them is Microsoft's

This note is not about a wrong number on somebody else's website. It is about where a number comes from, which turns out to decide almost everything about whether it can be trusted six months later.

Here is the case that prompted it. A widely cited AI pricing tracker — not named here, because the point is structural and naming it would make it look personal — states that Azure OpenAI's Global and Regional deployment types carry the same per-token rate, and lists one price per model. Microsoft's own documentation states the opposite, in a sentence with numbers in it.

The two sentences

Pricing varies by deployment type: Global (lowest) → Data Zone (~10% premium) → Regional (10–25% premium over Global, varying by region).

Microsoft Learn, Azure OpenAI in Azure AI Foundry Models FAQ — learn.microsoft.com/azure/ai-foundry/openai/faq, retrieved 7 August 2026

Against that, the tracker's position is the absence of a distinction: one rate per model, no deployment-type dimension at all, so a reader planning a Regional deployment in Europe reads a Global price and calls it their budget. On a modest production workload that is a quiet fifth of the bill, and it is quiet because nothing about the page looks uncertain.

Neither sentence is evidence. The meter is.

A vendor FAQ is still a secondary source about the vendor's own prices, so we did not stop there either. Microsoft publishes the actual billing meters through a public, unauthenticated retail-price API. We queried it for a single model and read the rates back per region and per deployment type.

The meter confirms the direction and the mechanism. It also shows why a quoted range is not a substitute for a reading: the premium is 10 percent at the bottom, and 32 percent in North Europe on the day we looked, outside the 10 to 25 percent the FAQ describes. We do not read that as the documentation being wrong — a published range can be typical rather than absolute, and one region on one day cannot settle which it is. It is simply the reason to read the meter even when a secondary source agrees with you: the range tells you roughly what to expect, the meter tells you what you will actually be charged. The tracker, meanwhile, was not slightly off. It was missing the dimension the price varies along.

Why this happens by construction

Aggregated pricing data is compiled once and then cited. The citation carries no memory of how it was compiled, so an error introduced upstream — a simplification, a missed deployment type, a rate that was true last quarter — propagates downstream in a form that looks exactly like a verified fact. Nobody lies. The error simply survives, because there is no step anywhere in the chain whose job is to go back and look.

A connector that reads live vendor data does not have that failure mode, and not because it is more careful. It has no upstream to inherit from. The rate it applies is the rate the vendor is publishing at the moment of the call, and when the vendor changes it, the next read changes with it. A compiled table cannot do that at any level of diligence, because being current is not a property of the diligence, it is a property of the pipeline.

This is the same discipline we hold ourselves to, and it is why these notes look the way they do. Note 1 reproduced the captured response envelope rather than describing it. Note 2 printed the engine's real output, minus sign included, rather than paraphrasing a verdict. The rule is the same one applied outward here: show the artifact, and if there is no artifact, do not make the claim.

Update, 7 August 2026: it is not an Azure quirk

After publishing we checked whether the deployment-region dimension exists anywhere else, on the assumption that a structure this easy to miss is rarely unique. It is not. Google's Vertex AI price list splits its Claude catalog into two explicitly named groups — "Models with regional pricing" and "Models with uniform pricing across all regions" — so whether the region dimension exists at all depends on which model you pick from the same vendor on the same page.

Two vendors, two different shapes: Azure's premium varies by region, Google's is a flat ten percent, and Google applies it to some models and not others. What they share is the part that matters here — a headline rate that is the cheapest of several, and a reader who has no way to tell from a one-price-per-model table that they were quoted the floor rather than their rate. We hold no per-region prices ourselves today; every rate in our own catalog is a global rate, and we would rather say that than imply a coverage we do not have.

The practical version, for anyone pricing an AI workload: whichever tool you use, ask where the number came from. If the answer is another tool, you are one link further from the meter than you think.