routerllm

Smart LLM request router — automatically route prompts to the right model based on cost, speed, quality or complexity.

pip install routerllm

The Problem

You have access to multiple LLMs — cheap ones, fast ones, powerful ones. But you use the same model for everything. Simple questions go to GPT-4o (overkill, expensive). Complex analysis goes to GPT-4o-mini (not enough, bad output).

routerllm fixes this automatically.

Quick Start

from routerllm import Router

router = Router(strategy="complexity")

router.add("openai", "gpt-4o-mini", api_key="sk-...")  # cheap, fast
router.add("openai", "gpt-4o", api_key="sk-...")       # powerful, expensive

# Simple → gpt-4o-mini
result = router.complete("What is 2+2?")
print(result.model)           # "gpt-4o-mini"
print(result.cost_usd)        # ~$0.000002

# Complex → gpt-4o
result = router.complete(
    "Analyze the long-term macroeconomic impact of AI on developing nations, "
    "comparing Keynesian and neoclassical perspectives."
)
print(result.model)           # "gpt-4o"
print(result.routing.reason)  # "Complex prompt (score: 0.78) → most powerful model"

Strategies

# Set default strategy
router = Router(strategy="cost")

# Override per call
result = router.complete("Explain quantum entanglement", strategy="quality")

Dry Run (no API call)

decision = router.dry_run("What is machine learning?")
print(decision)
# {
#   "would_use": "openai/gpt-4o-mini",
#   "strategy": "complexity",
#   "reason": "Simple prompt (score: 0.28) → cheapest model",
#   "complexity_score": 0.28,
#   "estimated_cost_per_1k": 0.00075
# }

Multi-Provider

router = Router(strategy="complexity")

router.add("openai", "gpt-4o-mini", api_key="sk-...")
router.add("openai", "gpt-4o", api_key="sk-...")
router.add("anthropic", "claude-haiku-4", api_key="sk-ant-...")
router.add("anthropic", "claude-opus-4", api_key="sk-ant-...")
router.add("gemini", "gemini-2.5-flash", api_key="AIza...")

RouterResult

result.output           # LLM response text
result.model            # Model that was used
result.provider         # Provider that was used
result.cost_usd         # Cost in USD
result.tokens_in        # Input tokens
result.tokens_out       # Output tokens
result.latency_ms       # Response time in ms
result.success          # True if no error
result.routing.reason   # Why this model was chosen
result.routing.complexity_score  # 0.0-1.0

License

MIT