Switchboard Source

The right Claude model for every step.

Switchboard plans your task, sends each step to the cheapest model that can do it well, checks the work, and switches models mid-task: up when a check fails, back down when the work gets easy.

Run it live with your own API key

Your key stays in this browser and is sent only to api.anthropic.com. Without a key you can still preview the routing for free.

Benchmark

Simulated · no API calls

Cost by kind of prompt

Switchboard vs one Opus call

By routing mode

How it's measured: 60 prompts in 6 categories (bench/prompts.json). Switchboard's free rules route every step. Token counts come from a cost model, not real responses. The baseline is one Opus 5.5 call writing the same amount of output. Switchboard's side includes the triage, planning and check calls. Thinking tokens aren't counted on either side. Reproduce it with npm run bench.

How it works

TriageFree rules score the prompt. Haiku breaks ties.
→
PlanBig tasks get split into steps, each with its own difficulty.
→
RouteEach step goes to the cheapest model for its difficulty.
→
CheckHaiku grades the step. Fail → move up a model and retry.
→
ComposeFinished steps are stitched into one answer, with no extra call.
for step in plan(task):
    model = cheapest_model_for(step.difficulty)
    loop:
        out = run(model, step)          # busy or refused → next model
        if check(out).passes: break
        if over_budget or at_top: break
        model = one_level_up(model)     # escalate and retry
    after 2 clean passes: drop back down