The interesting number is not how many models. It is where each one processes your request.
One route to every model, and a content policy that route cannot travel without. The catalogue is produced from the provider's live list by a script, classified by where inference actually runs, and read by this website, the product and the agent index alike. Nobody retypes a model name into copy.
A script reads the provider's live catalogue, resolves what a deployment can actually reach, and labels every model by where a request is processed when the deployment sits in Canada. The classification is the product of the exercise, not the count.
Generated on a date, so it moves. Per-account entitlement varies, and refreshing it is part of managed operations.
The in-Canada generation set is previous-generation: Claude 3 Sonnet and Haiku, Llama 3 8B and 70B, Mistral Large 24.02, Mixtral 8x7B, Mistral 7B. Every current frontier model routes outside Canada. Data residency is configurable, pinned to your region when you need it, and it is a real choice with a real capability cost. The whole argument, with the arithmetic.
Your content policy is not a toggle in an administration screen. It is attached to every inference call, and cloud identity policy denies the call without it. Four denies sit on the chat role and on every module inference role.
Pin a deployment to a jurisdiction and the residency control becomes an allow scoped to in-country destinations plus a request-time refusal. If a model has no compliant route, the request is refused with a reason. It is never quietly sent somewhere else, and it is never downgraded to a lesser model on your behalf.
Under the default posture, prompts are not used to train models and are not shared with model vendors, per the managed model service's terms. The election is made at account scope, and the workload is denied from widening it.
One vendor offers its newest frontier tier only with provider-side data sharing and retention of up to 30 days. Those models render unavailable until you elect that explicitly. It never happens by upgrade.
| State | What it means | What happens |
|---|---|---|
| Serveable | Allowlisted, entitled, with a compliant route from your deployment | ✓Available in chat |
| Needs access | In the catalogue, but your account has not been granted it by the provider | △We request it; some vendors want a use-case form |
| Needs a retention election | Offered only with provider-side data sharing | △Shown unavailable until you elect it |
| No route in your region | Your deployment is pinned and the model has no compliant path | ✕Refused in words, with the reason |
| Ungoverned plane | Cannot carry the content policy as a deny | ✕Listed and classified, never selectable |
That last row is the discipline in one line. A second model plane exists and could be switched on tomorrow. It cannot carry your content policy as a deny in cloud identity policy, and it has no Canadian region at all, so on a governed deployment it is shown, classified and refused rather than quietly offered.
For an order of magnitude: a knowledge question of roughly 10,000 tokens in and 1,000 out costs about 4 to 5 cents at a mid-tier frontier model's list price. That is an estimate, not a quote, and it lands on your cloud invoice rather than ours. Compute costs.
You do, and we apply it by release. The allowlist is a deploy-time parameter checked at listing and again at invoke, which means no administrator, including ours, can widen it from inside the running system. See administration.
Under the default posture, prompts are not used to train models and are not shared with model vendors, per the managed model service's terms. The single exception is the newest frontier tier described above, which is unavailable unless you elect it.
We re-run the generation script as part of managed operations. The site, the ticker and the product's own model library read the same generated file, so a change shows up in all three or in none. That is why the date is printed next to the number.
You could, and it is a genuine alternative with a different bill and a different operating burden. The catalogue already includes open-weight families from several labs through the managed service. We set the trade out on self-hosted models.
Yes, on the Strict tier. Read the numbers at the top of this page first, then posture tiers. Most organizations find that storage in their own region plus a governed route is the decision they actually wanted.
If we retyped it, the copy would be right on the day it was written and wrong every day afterwards.
We will run the residency arithmetic against your actual obligations and tell you what it costs in capability before you decide.