Point a swarm of cheap models at a hard question. A strategic planner decomposes the work; volume runs on the cheapest capable model. You set the caps.
ROUTER ONLINE14 workers idle
01
Objective
plain language — the swarm figures out the rest
Try an example
02
Topology
how the agents are wired
SWARM
Parallel fan-out. One planner splits the goal into independent hypotheses; many cheap agents test them at once.
LOOP
Self-improve until converged. Each round critiques the last and rewrites — stops when the score plateaus or a cap is hit.
COMBO
Loop of swarms. Fan out, critique the whole cohort, refocus the next fan-out. Deepest search, highest spend.
03
Model routing
cheapest capable model, picked per task
Worker pool · auto-router$ / MTokshare
Claude Haikufast · cheap · default
$0.80 / $4.00
61%
Gemini Flashlong-context sweeps
$0.30 / $2.50
27%
Claude Sonnetescalation only
$3.00 / $15.0
12%
PICKED BY TASKRouter benchmarks each sub-task and sends it to the cheapest model that clears the bar. Sonnet only on adversarial verification.
◆ Planner: Fable 5
Strategic decomposition, hypothesis design, and final synthesis run on Fable 5 — the expensive model does the thinking, cheap models do the volume.
Planner budget cap$
≈ 8% of spend on Fable · 92% on cheap workers. Strategy from Fable, scale from the swarm.