We have a bit of a saying, "Welcome to the Cloud! A new experience every day!". Unfortunately this is particularly relevant to AI/ML, which seems to change even more rapidly. Please forgive us (and let us know) for anything that is out of date here!
¶ Model availability
See Deployment types below for a deeper dive, but our straightforward reccomendations depend on your experiment type:
¶ Unclassified experiments
If you're running an unclassified experiment (i.e., not Protected B) experiment, you can actually use models from any region. When new OpenAI models are deployed, they tend to be first available in the US; as of Oct '26 however, both Canada Central and East have all the latest models, until newer models (than GPT-6.1) are released, Global Standard in either Canadian region should be fine.
See here for the region table for Global Standard model availability.
¶ Protected B experiments
Protected B data must be stored in Canada, although it can be processed outside of Canada (so long it remains encrypted in transit).
Although TBS policy does not require PB to be processed domestically, this does not mean your project's ATO/SA&A1 will accept this risk. As a result, you may want to plan for the very decent chance you'll be required to keep everything, including processing, in Canada.
1ATO: Authority to operate, SA&A: Security Audit & Authorization. See here for a backgrounder on ATO/SA&As.
In practice, with Azure OpenAI, this means you need to deploy your resources in Canada (specifically Canada East for the best options).
In our sandbox space, we do permit processing outside of Canada by using Global deployments.
Please still use your judgement, if processing your PB outside of Canada makes sense!
If you want to process in Canada, Standard deployment model availability lists the options (as of October 2025):
- gpt-4o, version 2024-11-20
- gpt-35-turbo, version 1106 or 0125
If you decide to process outside of Canada, you can use models per Global Standard model availability.
¶ Moving to production
In our experience, almost all cloud prod spaces are setup to support Protected B, meaning even if your project does not use PB, you should still expect the restrictions of PB (see Protected B experiments above).
Note that in production you may also have the option of Provisioned deployments, which are essentially where you make a commitment of tokens/hour, but there are a couple more models offered.
¶ Deployment types
In Azure, there is primarily two 'dimensions' to how you deploy models, the geographic and type. See below, or MS' documentation.
¶ Geographic
- Standard or Regional models are processed in the region of the (OpenAI or AI Foundry) resource they're attached to.
- Global, as the name implies, more or less ignores the region of your resource and will process your request at whatever Azure datacenter that can serve you best. Strangely however, you cann't deploy all models as Global in all regions; see the Unclassified experiments section above for our reccomendation.
- Data Zone is a special one that doesn't really apply to us; it's like Global but limits processing to the same geopolitical boundary the resource is deployed in: either the US, EU, or APAC.
¶ Type
- Standard is the baseline option.
- Provisioned or Provisioned Managed are more or less like a commitment tier, where you commit to some hourly number of tokens (using an absurdly complicated PTU metric). GPT-4o in this deployment costs CAD$277/h for 50M tokens. Since this is pretty unlikely to be appropriately used in our environments, we've used Azure Policy to prevent accidental deployment of this type; they can be safely ignored.
- Batch is, unsuprisingly, for batched requests. Although you do need to adjust your API calls / application logic to support it, batch can save a lot of cost, as MS seems to price it at 50% of the equivalent non-batch price.
- Managed Compute (aka Hugging Face models) is a newer option that gives you access to a 'curated' selection of open weight models. Pricing, while reasonable for what you're getting, is unfortunately similar to Provisioned above: you pay for the allocated hardware regardless of how much or little you use it and charges accrue until the deployment is removed. Since there is significant utility/interest, we've left this option available; please be aware of the costs and remove your deployments as soon as they're no longer required! In the last step before deploying one of these models, there will be a confirmation indicating the hourly cost.
The above geographic & type options are combined to form the ~8 different possible ways to deploy a model; unfortunately the use of the word 'Standard' could imply either the geographic or type, and is sometimes entirely omitted 
.
¶ Pricing
Most Azure OpenAI pricing is by tokens, which of course are an AI/ML way of representing (usually) text. As a rule of thumb, one token tends to be equal to about 4.5 characters; so short, common words are usually one token, with longer or more obscure jargon being more. There are tools out there like TikToken that can measure content without having to send it to OpenAI, if that's something you need.
¶ Actual costs
Azure OpenAI pricing is the authoritative resource.
- Don't forget to set your region & currency!
- Make sure you're looking at exactly the right model;
o4,4o, and4o-realtimeare all very different! - Note the different type of tokens:
- Input are tokens sent to the model
- Cached (certain models only) are tokens that have already been sent to the model in a previous turn (i.e. followup questions. For models that don't support caching, all tokens, including those previously prompted, are counted as input)
- Output are the ouputs from the model. As a rule of thumb, cost 4× more than Input tokens.
- Batch does require a different API/app logic, but costs about 50% vs realtime.