DCAIAPI reference#
Base URL and authentication#
| API | Base URL | Key header |
|---|---|---|
| OpenAI v1 and Azure OpenAI | https://ai-gateway-east-us1.azure-api.net/gateway |
api-key: <gateway key> |
| Anthropic Messages | https://ai-gateway-east-us1.azure-api.net/gateway/anthropic |
x-api-key: <gateway key> |
Keep the key in a header. The gateway also accepts it as the subscription-key query parameter (it removes the key before
forwarding the request and hides it in its own logs), but URLs tend to end up in logs on your side, such as proxies and
browser history. The OpenAI SDKs send their api_key as Authorization: Bearer, so add default_headers={"api-key": key} as the
samples do. The Azure OpenAI and Anthropic SDKs already use the right header.
Endpoints#
| Method and path | Speaks | Notes |
|---|---|---|
POST /openai/v1/chat/completions |
OpenAI Chat Completions | Streaming supported |
POST /openai/v1/responses |
OpenAI Responses | |
POST /openai/v1/embeddings |
OpenAI Embeddings | |
POST /openai/deployments/{deployment}/chat/completions?api-version=… |
Azure OpenAI | For the Azure OpenAI SDKs; translated to the v1 call above |
POST /openai/deployments/{deployment}/embeddings?api-version=… |
Azure OpenAI | Same |
POST /anthropic/v1/messages |
Anthropic Messages | Streaming supported |
POST /anthropic/v1/messages/count_tokens |
Anthropic token counting |
Request and response bodies are those of the API you call; the gateway passes them through. It changes only the model
field, to the deployment it chose.
How the gateway chooses a deployment#
For the OpenAI-compatible endpoints, in this order:
x-dca-routeheader present: the deployment you mapped to that route on the Connect your Foundry page. The request'smodelis ignored. An unmapped route uses your default deployment, exceptoss-only, which is refused.oss-onlyis a guarantee, not a label: it is also refused (422 not_open_weight) when the deployment mapped to it is not billed on the open-weight tier (see Limits and billing).cheapis the deployment you chose for that route, not an automatic price comparison.- No route,
modelnames a deployment: that deployment of your resource. Any deployment of the connected resource can be named. - No route, no
model: your default deployment.
Embeddings use the deployment mapped to embeddings, else the one named in model. The Anthropic API uses the deployment
mapped to Claude, else the one named in model.
Routes#
x-dca-route |
Intended for | If you did not map it |
|---|---|---|
cheap |
The lowest-cost deployment you chose (a mapping, not an automatic price comparison) | Default deployment |
oss-only |
Open-weight models only; enforced: the mapped deployment must bill as open-weight, else 422 not_open_weight |
422 route_not_configured; never sent to the default |
auto |
A model-router deployment that picks per request | Default deployment |
reasoning |
A reasoning model | Default deployment |
code |
A coding model | Default deployment |
frontier |
Your most capable model | Default deployment |
passthrough or no header |
Use model, else the default |
Data boundary#
x-dca-data-boundary |
Behaviour |
|---|---|
any (default) |
Uses the deployments in your main mapping. Where inference runs depends on each deployment's type (Global, Data Zone or regional). |
us |
Uses only the deployments you declared for the United States data boundary. A request with no declared deployment is refused with 422 no_eligible_provider and is never rerouted. |
| anything else | 400 invalid_data_boundary |
The gateway enforces your declaration. It cannot see your deployments' types, so declare only US data-zone or US regional deployments.
Request headers#
| Header | Values | |
|---|---|---|
api-key / x-api-key |
Your gateway key | Required |
x-dca-route |
See Routes | Optional |
x-dca-data-boundary |
any, us |
Optional |
Response headers#
| Header | Example | Meaning |
|---|---|---|
x-dca-backend |
customer-foundry |
Your own Foundry served the request |
x-dca-model |
gpt-4.1-mini |
The deployment the gateway called |
x-dca-data-boundary |
any |
The boundary applied |
x-dca-processing-scope |
customer-deployment |
Processing happened in your deployment, under its own region or data zone |
x-dca-remaining-tokens |
59870 |
Tokens left in the current minute for this key |
x-dca-remaining-quota-tokens |
499999824 |
Tokens left this calendar month for this key |
x-dca-budget-remaining-tokens |
99999976 |
Tokens left this calendar month in the budget you set for this key (only on keys with a budget) |
x-dca-tokens-consumed |
24 |
Prompt and completion tokens of this request |
retry-after |
12 |
Seconds to wait, on 429 and 403 limit responses |
Your deployment's own headers, for example its rate-limit headers, pass through as well.
Streaming#
"stream": true works on chat completions and the Anthropic Messages API; events pass through as they arrive. With the
OpenAI API, add "stream_options": {"include_usage": true} if your application needs the usage totals in the stream.
Compatibility notes#
- OpenAI SDKs: set the base URL to
…/gateway/openai/v1and add theapi-keydefault header. - Azure OpenAI SDKs: set the endpoint to
…/gateway;modelis your deployment name. Use API version2024-10-21or later. - Anthropic SDKs: set the base URL to
…/gateway/anthropic. Theanthropic-versionheader defaults to2023-06-01if you omit it. - Timeouts: the gateway forwards while your deployment generates. Use your SDK's normal timeouts for long generations.