Skip to main content

​
Router

The routing layer. Selects the right model for each call based on rules, context budget, and cost constraints. Runs entirely in-process.

​
Why routing matters

Naive routing — “short prompt = cheap model” — silently breaks things:
  • A 150k token conversation sent to a model with an 8k context window gets truncated without warning
  • Switching from Sonnet to Haiku for a complex multi-constraint prompt causes capability regression
  • There is no way to know routing failed until a user reports a wrong answer
TokenSense’s router checks the context budget before routing and escalates automatically on failure.

​
Basic setup

from tokensense.router import Router, Rule

router = Router(
    tiers={
        "small":  ["claude-haiku-4-5"],
        "large":  ["claude-sonnet-4-6"],
    },
    rules=[
        Rule(if_context_tokens_gt=4000, deny_tiers=["small"]),
        Rule(if_task="legal-review", pin_tier="large"),
    ],
    on_failure="escalate",
)

​
Getting a routing decision

messages = [{"role": "user", "content": "Summarise this contract..."}]

decision = router.route(
    messages=messages,
    task_hint="legal-review",
)

print(decision.model)   # claude-sonnet-4-6
print(decision.tier)    # large
print(decision.reason)  # pinned to large by rule

​
Making the call with the routed model

client = observe(anthropic.Anthropic())

decision = router.route(messages=messages, task_hint="summarise")

response = client.messages.create(
    model=decision.model,
    max_tokens=1024,
    messages=messages,
)

​
Tiers

A tier is a named group of models. Models within a tier are tried in order — the first one whose context window fits the conversation is selected. 
tiers={
    "small":  ["claude-haiku-4-5", "groq/llama3-8b"],  # tried left to right
    "medium": ["gpt-4o-mini"],
    "large":  ["claude-sonnet-4-6", "gpt-4o"],
}
Tier order matters. The router tries tiers from first to last in the dict. Earlier tiers are preferred unless rules say otherwise.