Smart routing

Use ninja/auto to select a model for your request. Set routing.strategy to prioritize cost, latency, quality or a balanced approach. NinjaChat selects an available provider and can try another route if the request fails. GET /api/v1/models/ninja/auto lists the router’s current candidate pool under router.candidates; the per-request ranking comes back on every response under routing.router.

const response = await client.responses.create({
  model: "ninja/auto",
  input: "Review this repository migration plan.",
  routing: {
    strategy: "balanced",
    allow_fallbacks: true,
    data_policy: "zero_retention",
  },
});

console.log(response.routing.resolved_model);
console.log(response.provider, response.cost_usd);

response = client.responses.create(
    model="ninja/auto",
    input="Review this repository migration plan.",
    routing={
        "strategy": "balanced",
        "allow_fallbacks": True,
        "data_policy": "zero_retention",
    },
)

print(response["routing"]["resolved_model"])
print(response["provider"], response["cost_usd"])

curl https://www.ninjachat.ai/api/v1/responses \
  -H "Authorization: Bearer $NINJACHAT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model":"ninja/auto",
    "input":"Review this repository migration plan.",
    "routing":{"strategy":"balanced","allow_fallbacks":true,"data_policy":"zero_retention"}
  }'

Routing strategies

With ninja/auto the strategy picks the model; with a pinned model it orders the provider rails behind that model.

Strategy	With `ninja/auto` (model choice)	With a pinned model (rail choice)	Reach for it when
`balanced`	Curated order for the detected task	Overall production fit	Default application traffic
`cost`	Cheapest model that clears the task’s quality bar, by reference-request price	Lowest eligible rail cost	High-volume extraction and classification
`latency`	Lowest p50 latency measured from live traffic	Lowest observed latency	Interactive chat and tool loops
`quality`	Flagship first	Highest-quality eligible rail	Final answers and complex agent work

Billing: you pay the resolved model and provider rail’s published rate for actual token usage. The response identifies both, so cost attribution stays explicit. See Pricing for the live catalog.

Provider controls

routing.providers lets you include, exclude, or order provider rails without changing the public model ID. data_policy filters rails before ranking, and require_parameters rejects rails that cannot honor every requested feature.

TypeScript SDK

const response = await client.responses.create({
  model: "claude-sonnet-5",
  input: "Draft a production rollout plan.",
  routing: {
    strategy: "latency",
    providers: {
      only: ["anthropic", "deepinfra", "gmicloud"],
      order: ["anthropic", "deepinfra"],
    },
    allow_fallbacks: true,
    require_parameters: true,
    max_cost_usd: 0.08,
  },
});

The response always tells you what ran:

{
  "model": "claude-sonnet-5",
  "provider": "anthropic",
  "routing": {
    "strategy": "latency",
    "requested_models": ["claude-sonnet-5"],
    "resolved_model": "claude-sonnet-5",
    "provider": "anthropic",
    "fallbacks_allowed": true,
    "data_policy": "default"
  }
}

When the request was ninja/auto, model and routing.resolved_model carry the concrete model that served it, routing.requested_models keeps ["ninja/auto"], and a router block explains the decision:

{
  "model": "gemini-3-flash",
  "provider": "google",
  "routing": {
    "strategy": "balanced",
    "requested_models": ["ninja/auto"],
    "resolved_model": "gemini-3-flash",
    "provider": "google",
    "fallbacks_allowed": true,
    "data_policy": "default",
    "router": {
      "id": "ninja/auto",
      "task": "summarize",
      "classified_by": "regex",
      "override": null,
      "reasoning": "...",
      "candidates": ["gemini-3-flash", "..."],
      "considered": ["..."]
    }
  }
}

task is one of the task types above, classified_by is regex, llm, or llm-cached, candidates is the ranked chain the router would try in order, and considered lists every model it evaluated with the reason it was kept or eliminated. The same block is persisted on the request trace (GET /requests/{id}).

Current frontier choices

The router tracks the live catalog rather than a hard-coded marketing table. Current leading options you can pin directly include gpt-5.6-sol, gpt-5.5-pro, claude-fable-5, claude-opus-5, claude-sonnet-5, gemini-3.7-flash, gemini-3.1-pro, grok-4.6, deepseek-v4-pro, glm-5.2, minimax-m2.7, and qwen3-coder-next. Fetch /models at runtime when you need a user-facing picker.

Pair it with

Spend controls

routing.max_cost_usd caps each request

Fallbacks

Explicit order with automatic failover

Quality gates

Evaluate important outputs before acceptance

Get started

Controls

Routing strategies

Provider controls

Current frontier choices

Pair it with

Spend controls

Fallbacks

Quality gates

​Routing strategies

​Provider controls

​Current frontier choices

​Pair it with

Spend controls

Fallbacks

Quality gates

Routing strategies

Provider controls

Current frontier choices

Pair it with