llm-router

Route prompts to the right LLM automatically by complexity. Supports Groq, OpenAI, and Anthropic. Cut costs without cutting quality.

Stop hardcoding a single model. llm-router scores your prompt's complexity and routes it to the cheapest model that can handle it.

Simple prompts → fast cheap models. Complex prompts → powerful models. Automatically.

---

Package name note: this repo is llm-router but the PyPI package is llm-dispatch.

pip install llm-dispatch   # ← correct
pip install llm-router     # ← does NOT exist on PyPI

The import is from llmrouter import LLMRouter.

---

When to use this

Use llm-router when:

• You're making many LLM calls with varying complexity and want to minimize cost automatically
• You want to run simple queries on cheap/fast models (Groq Llama 8b) and only escalate to GPT-4o or Claude Sonnet when the prompt actually needs it
• You support multiple providers (Groq + OpenAI + Anthropic) and want a single interface
• You want per-call cost estimates logged alongside responses

Not the right fit if you need agents, tool calling, or function execution — this is a routing layer for text completion, not an agent framework.

---

Why not LiteLLM?

LiteLLM is a unified API layer — it lets you call any model with the same interface, but you still decide which model to use. llm-router decides for you based on prompt complexity. The two tools are complementary, not competing: use LiteLLM if you want provider unification with manual model selection; use llm-router if you want the model selection to be automatic and cost-aware.

---

Install

pip install llm-dispatch

# Install provider SDKs you need:
pip install llm-dispatch[groq]       # Groq only
pip install llm-dispatch[openai]     # OpenAI only
pip install llm-dispatch[anthropic]  # Anthropic only
pip install llm-dispatch[all]        # All providers

---

Quick Start

from llmrouter import LLMRouter

router = (
    LLMRouter(verbose=True)
    .add("groq/llama-3.1-8b-instant")    # fast, cheap
    .add("groq/llama-3.3-70b-versatile") # powerful
    .add("openai/gpt-4o")                # best quality
)

# Simple prompt → routed to cheap fast model
result = router.complete("What is the capital of France?")
print(result.output)      # Paris
print(result.model_used)  # llama-3.1-8b-instant
print(result.estimated_cost_usd)  # 0.000001

# Complex prompt → routed to powerful model
result = router.complete("""
    Analyze the architectural trade-offs between microservices and
    monolithic systems for a high-traffic fintech application.
    Consider scalability, fault tolerance, and deployment complexity.
""")
print(result.model_used)  # gpt-4o

---

Strategies

# Auto (default) — complexity-aware routing
router = LLMRouter(strategy="auto")

# Always cheapest model that fits the context window
router = LLMRouter(strategy="cheapest")

# Always fastest model
router = LLMRouter(strategy="fastest")

# Always highest quality model
router = LLMRouter(strategy="smartest")

# Balance speed and quality
router = LLMRouter(strategy="balanced")

# Override per call
result = router.complete(prompt, strategy="cheapest")

---

How auto-routing works

The router scores each prompt from 0.0 (trivial) to 1.0 (very complex):

Complexity	Score	Routed to
Simple question	< 0.25	Cheapest fast model
Medium task	0.25–0.60	Best quality/cost ratio
Complex analysis	> 0.60	Highest quality model

Scoring factors: prompt length, technical keywords, code blocks, number of questions, multi-part instructions.

from llmrouter import complexity_score

complexity_score("What is 2+2?")                         # 0.0
complexity_score("Summarize this 500-word article")      # 0.35
complexity_score("Refactor this Python class and add type hints and unit tests") # 0.72

---

Available Models

from llmrouter import PRESET_MODELS
print(list(PRESET_MODELS.keys()))

Model ID	Provider	Speed	Quality	Cost/1k tokens
groq/llama-3.1-8b-instant	Groq	fast	4/10	$0.00005
groq/llama-3.3-70b-versatile	Groq	medium	8/10	$0.00059
groq/mixtral-8x7b-32768	Groq	medium	7/10	$0.00024
openai/gpt-4o-mini	OpenAI	fast	6/10	$0.00015
openai/gpt-4o	OpenAI	medium	9/10	$0.005
anthropic/claude-haiku-3	Anthropic	fast	6/10	$0.00025
anthropic/claude-sonnet-4	Anthropic	medium	9/10	$0.003

---

Custom Models

from llmrouter import LLMRouter, ModelConfig

router = LLMRouter()
router.add_custom(ModelConfig(
    model_id="my-provider/my-model",
    provider="groq",           # uses Groq's SDK
    name="my-model-name",
    cost_per_1k=0.0001,
    context_window=16384,
    speed="fast",
    quality=5,
))

---

RouteResult

Every .complete() call returns a RouteResult:

result.output               # str — model response
result.model_used           # str — model name selected
result.provider             # str — 'groq' | 'openai' | 'anthropic'
result.complexity_score     # float 0.0–1.0
result.strategy             # str — strategy used
result.estimated_cost_usd   # float — estimated cost in USD

---

Environment Variables

GROQ_API_KEY=gsk_...
OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...

Or pass per model:

router.add("groq/llama-3.3-70b-versatile", api_key="gsk_...")

---

Author

M. Adhitya — Founder, Rewrite Labs

License

MIT © 2025 M. Adhitya

Built at Rewrite Labs