WGRIN Tools · Azure AI models
About the Azure AI model catalogue
How WGRIN collects the data, what it means and what it cannot tell you.
Sources
- Availability: Microsoft's model catalogue API (Azure Resource Manager Models - List, per region), read by WGRIN's backend with a read-only identity. It lists what Microsoft offers in a region; the page says "as listed by Microsoft's model catalogue API on <date>". Another subscription or offer type may see a different list.
- Prices: the public Azure Retail Prices API, in USD (Microsoft's authoritative retail currency). They are list prices without discounts (EA, MCA, CSP, savings plans, reservations).
- Context window and output limits: Microsoft Learn, recorded by WGRIN with the source and document date, and shown as "documented", not as an API fact.
- Location facts: Microsoft Learn: deployment types.
Microsoft's APIs are called only by WGRIN's weekly background refresh (or an administrator's refresh), never by your visit. Model names, versions and prices are matched by explicit, reviewed rules; when a price cannot be matched without guessing it stays unknown. Prices are never filled in by analogy (for example "Data Zone = Global + 10 %").
Region, resource and where data is processed
Four separate facts, which are easy to confuse:
- Selected catalogue region (Sweden Central): the region whose Microsoft catalogue listing and retail prices are shown.
- Resource region: where you would create the Azure AI Foundry resource and deployment. It is the same code, but it is a statement about your resource, not about where requests are processed.
- Inference processing scope: depends on the deployment type, see the table below.
- Data at rest: stays in the resource's geography for all deployment types.
| Deployment type | Inference processing | Data at rest |
|---|---|---|
| Data Zone Batch (DataZoneBatch) | Within the EU data zone (Microsoft's EU Data Boundary, which may include EFTA countries such as Norway and Switzerland), not only in the selected region. | Stays in the resource's geography. |
| Data Zone Provisioned (DataZoneProvisionedManaged) | Within the EU data zone (Microsoft's EU Data Boundary, which may include EFTA countries such as Norway and Switzerland), not only in the selected region. | Stays in the resource's geography. |
| Data Zone Standard (DataZoneStandard) | Within the EU data zone (Microsoft's EU Data Boundary, which may include EFTA countries such as Norway and Switzerland), not only in the selected region. | Stays in the resource's geography. |
| Developer (DeveloperTier) | No data residency guarantee. | Stays in the resource's geography. |
| Global Batch (GlobalBatch) | Any Azure region worldwide (Microsoft's global infrastructure). | Stays in the resource's geography. |
| Global Provisioned (GlobalProvisionedManaged) | Any Azure region worldwide (Microsoft's global infrastructure). | Stays in the resource's geography. |
| Global Standard (GlobalStandard) | Any Azure region worldwide (Microsoft's global infrastructure). | Stays in the resource's geography. |
| Regional Provisioned (ProvisionedManaged) | Within the resource's geography. | Stays in the resource's geography. |
| Standard (regional) (Standard) | Within the resource's geography. | Stays in the resource's geography. |
No deployment type is described here as processing data only in Sweden Central: Data Zone deployments may process requests anywhere in the data zone, Global deployments anywhere in the world.
Token counts and exactness
The calculator counts tokens in your browser with the model's encoding (for example o200k_base). Each model has an
exactness level: Verified (an official source names the encoding for that model), Inferred (only an official prefix rule
or a statement about a predecessor, for example GPT-5.6 via tiktoken's gpt-5 prefix) or Estimate (no evidence, for
example GPT-6). Inferred and Estimate counts are always shown as estimates. Counts cover the text you paste only: message framing,
system prompts, tool definitions, images and hidden reasoning tokens (billed as output) are not included. Browsers turn pasted line
breaks into LF, which is what most API clients send.
Privacy
Text you paste into the calculator is counted in your browser and is not sent to an AI model nor to WGRIN servers: no form submission, no network request, no storage, no analytics. Browser extensions and your browser's own cloud spell-check features are outside our control; the text area turns spell-check off.
Known limitations
- Availability comes from a subscription-scoped API; Microsoft offers no anonymous, global list.
- The Retail Prices API has no structured model, version or deployment type fields; matching uses reviewed rules, and a new meter family stays unknown until it is reviewed.
- Prices are mostly published per model, not per version.
- There is no Microsoft price history before WGRIN's first observation (2026-10-04).
- Short/long-context thresholds are not machine-readable; without a sourced threshold both prices are shown.
- Capability flags are incomplete: absence does not mean "not supported".
- Provisioned (PTU) costs, fine-tuning, batch sizing and cache-write costs are not modelled by the calculator.
- The refresh runs weekly, so a new model can appear up to a week late.