Limits and billing#

Per-key limits#

Each gateway key has two limits. They protect your Foundry and your bill from runaway clients. Every key of a subscription has its own counters, so one busy tenant cannot use up another tenant's allowance.

• Over the per-minute allowance, the gateway answers 429 with retry-after. Prompt tokens are estimated before the request is forwarded, so an oversized prompt is refused without reaching your deployment.
• Over the monthly cap, the gateway answers 403 until the next calendar month (UTC).
• Every response reports what is left: x-dca-remaining-tokens for this minute and x-dca-remaining-quota-tokens for this month.

Your own monthly budget per key. On the Tenant keys page you can give any key a lower monthly budget (at least 1,000 tokens), for example the amount a customer pays you for. Over its budget the key is refused with 403 until the next calendar month (UTC), while your other keys keep working. Responses on a key with a budget also carry x-dca-budget-remaining-tokens. The plan's monthly cap above still applies.

Your deployments keep their own quotas in Azure as well. Whichever limit is lower applies first.

Keys per subscription#

Revoked keys do not count towards the number of keys.

What you pay#

You receive two separate charges:

1. Model usage is billed by Microsoft on your own Azure subscription, as with any direct call to your Foundry deployments. DCAI Model Gateway does not resell or mark up model access.
2. DCAI Model Gateway is billed through Microsoft Marketplace on the same Azure invoice: a monthly plan fee, with routed tokens beyond the included amount metered per million.

Prices as listed on 2026-10-15. The Marketplace listing is authoritative. Taxes and currency follow your Azure billing account.

Routed tokens are the prompt and completion tokens of every request that reached your deployment. Refused requests (4xx and 5xx from the gateway) are not counted.

Tier. A request counts as open-weight when its deployment serves an open-weight model family (for example Llama, DeepSeek, gpt-oss, Mistral, Phi, Qwen, Gemma); everything else, including GPT and Claude models, counts as frontier. The gateway takes the model name each deployment reported in your last connection test, shown in the Billing tier column. A deployment the test has not called counts as frontier, so map and test every deployment you use (Save and test on the Connect page). For Google Vertex AI the model id itself decides: Gemini and Claude are frontier; open-weight models such as gpt-oss, DeepSeek, Qwen, GLM, Llama or Gemma are open-weight. Only a valid answer in the connection test sets a tier: a throttled (429) or malformed answer leaves it unconfirmed (frontier) until a later test succeeds, and a tier never carries over when a deployment name is pointed at another endpoint or project. The oss-only route is enforced against this tier: it is refused (422 not_open_weight) unless the mapped deployment bills as open-weight.

Active tenants. An active tenant is a key, the Default key included, that routed at least one token in the subscription's monthly billing period. Keys that are never used are free. Beyond the plan's included number, each additional active tenant is charged once for that billing period, however much it routes. The Tenant keys page shows how many keys are active in the current period.

Where to see usage#

• Per request: the x-dca-tokens-consumed response header.
• Per key: the Tenant keys page shows each key's tokens for the month you choose, and Download usage (CSV) gives tokens per day, key, deployment, billing tier, route and data boundary, for rebilling your own customers.
• Per month: the Marketplace charges for DCAI Model Gateway in Azure Cost Management and on your invoice, and your deployments' own metrics in Azure Monitor.