Models

Eridian routes production inference across GPT and Gemini. The Models API lists what your project may call in its residency region. Application code should prefer "auto" unless a workflow is pinned for evaluation or examiner evidence.

Powered primarily by OpenAI and Gemini. You can also point Eridian at a URL for a provider you run yourself. Eridian does not operate those GPUs.

Catalog

Do not pin SKU strings from a provider console. The control plane maps gpt and gemini to the versions your order form authorizes. Version rolls are recorded on the audit event as eridian.route plus eridian.provider_revision.

Private, local, and bare-metal providers

Hosted production families are GPT (OpenAI) and Gemini (Google). You can attach a provider you operate: a local model server, a private provider, or a bare-metal endpoint. Set the customer-configured base URL on the project. Inference goes to that URL. Eridian does not operate those GPUs.

{
  "provider_id": "desk_private_llm",
  "family": "private",
  "base_url": "https://models.customer.example/v1"
}

family is private. Do not send provider SKU strings. The Models API lists the private hop when the project has a URL configured. Health, PII, caching, and the audit trail still run in Eridian; the model weights stay on the endpoint you host.

List Models

GET /v1/models
Authorization: Bearer eridian_sk_...
X-Eridian-Project: prj_legal_001

Required scope: models:read

Example Response

{
  "object": "list",
  "data": [
    {
      "id": "gpt",
      "object": "eridian.model",
      "family": "gpt",
      "context_window": 128000,
      "supports_streaming": true,
      "supports_structured_output": true,
      "supports_vision": true,
      "regions": ["us-east-1", "eu-west-1", "me-central-1", "ap-southeast-1"],
      "status": "active"
    },
    {
      "id": "gemini",
      "object": "eridian.model",
      "family": "gemini",
      "context_window": 200000,
      "supports_streaming": true,
      "supports_structured_output": true,
      "supports_vision": true,
      "regions": ["us-east-1", "eu-west-1", "me-central-1", "ap-southeast-1"],
      "status": "active"
    }
  ]
}

Retrieve Model

GET /v1/models/gpt
Authorization: Bearer eridian_sk_...

Returns capability flags, context window, and current availability for the requested region. Unknown IDs return HTTP 404 with error.code: model_not_found.

Model Availability

Availability varies by data residency region and enterprise tier. Pin or restrict the pool via routing policies.

When a family is temporarily unavailable due to provider health, it is removed from the active pool until recovery. Requests that pin an unavailable family return HTTP 503 with error.code: model_unavailable.

Using "auto"

When model is "auto", the routing engine selects GPT, Gemini, or a configured private hop from your project's default routing policy. Override per-request with routing_policy:

{
  "model": "auto",
  "routing_policy": "latency_optimized",
  "project_id": "prj_dev_001"
}

Response field eridian.route is gpt, gemini, or private. Use it for debugging and cost attribution, not as a hard dependency in application branching unless the workflow is explicitly pinned.

Capability Matrix

Health

GET /v1/models/health
Authorization: Bearer eridian_sk_...

{
  "object": "eridian.model.health",
  "gpt": { "status": "healthy", "error_rate_5m": 0.002 },
  "gemini": { "status": "healthy", "error_rate_5m": 0.001 },
  "private": { "status": "healthy", "error_rate_5m": 0.0 }
}

See Multi-Model Routing for policy configuration and Inference for request fields.