Smart Routing

Stop choosing models. auto reads each prompt, detects the task, and routes it:

POST /api/v1/chat

{
  "model": "auto",
  "messages": [{"role": "user", "content": "Solve: x² + 5x + 6 = 0"}],
  "include_routing": true
}

Response

{
  "model": "o3-mini",
  "choices": [{"message": {"role": "assistant", "content": "x = -2 and x = -3"}}],
  "routing": {
    "requested": "auto",
    "resolved": "o3-mini",
    "task_type": "math",
    "reasoning": "Detected task type: math"
  },
  "cost": {"this_request": "$0.0008"}
}

include_routing: true shows which model was chosen and why.

Four variants

Model ID	Optimizes for	Reach for it when
`auto`	Quality + speed balance	You just don’t want to think about models
`auto-fast`	Lowest latency	Real-time chat, low-latency pipelines
`auto-cheap`	Lowest cost	High-volume jobs, classification, batch
`auto-quality`	Best output	Critical decisions, final drafts

Billing: you pay the resolved model’s metered rate for your actual token usage — route to o3-mini, you’re billed at o3-mini’s $/MTok; route to claude-sonnet-4.6, you’re billed at its (higher) rate. There’s no flat per-model price — see Pricing for the live rates. The routing field always names the cost-incurring model.

How detection works

By default (route_mode: "smart"), task classification runs through an LLM call — gemini-3-flash, cached in Redis for 1 hour and not charged to you — which is more accurate than regex for nuanced prompts. Regex-based signal matching (below) is the actual mechanism only for route_mode: "fast", plus auto-fast/auto-cheap (which skip the LLM step to stay low-latency), and it’s always the fallback if the LLM classifier errors.

In regex mode, the last 3 user messages are scanned for task signals:

Task	Signals	`auto` routes to
`code`	function, debug, implement, SQL, TypeScript…	`claude-sonnet-4.6`
`math`	equation, solve, integral, probability…	`o3-mini`
`creative`	write, story, poem, fiction, lyrics…	`claude-sonnet-4.6`
`analysis`	analyze, compare, evaluate, summarize…	`gpt-5`
`quick`	short prompts under 80 chars, “what is”…	`gemini-3-flash`
`general`	everything else	`gpt-5`

Full routing table — all four variants

Task	`auto`	`auto-fast`	`auto-cheap`	`auto-quality`
code	`claude-sonnet-4.6`	`claude-haiku-4.5`	`deepseek-v3`	`claude-opus-4.6`
math	`o3-mini`	`gpt-5-mini`	`gemini-2.5-flash`	`o3-mini`
creative	`claude-sonnet-4.6`	`gemini-3-flash`	`gemini-3-flash`	`claude-opus-4.6`
analysis	`gpt-5`	`gpt-5-mini`	`deepseek-v3`	`claude-opus-4.6`
quick	`gemini-3-flash`	`gemini-3-flash`	`gemini-2.5-flash`	`claude-opus-4.6`
general	`gpt-5`	`gpt-5-mini`	`llama-4-maverick`	`claude-opus-4.6`

In code

r = requests.post("https://www.ninjachat.ai/api/v1/chat",
    headers={"Authorization": f"Bearer {os.environ['NINJACHAT_API_KEY']}"},
    json={
        "model": "auto",
        "messages": [{"role": "user", "content": "Write a merge sort in Python"}],
        "include_routing": True,
    }
)
data = r.json()
print(data["choices"][0]["message"]["content"])
print("Routed to:", data["routing"]["resolved"])  # claude-sonnet-4.6

Pair it with

Budget Routing

budget_cents caps what auto may spend

Fallback Chains

Explicit order with automatic failover

Compare

Race models to pick one empirically

Get Started

Chat

Superpowers

Media & Search

Reference

Four variants

How detection works

In code

Pair it with

Budget Routing

Fallback Chains

Compare

​Four variants

​How detection works

​In code

​Pair it with

Budget Routing

Fallback Chains

Compare

Four variants

How detection works

In code

Pair it with