Delegation economics
Definition
A frontier lead owns the session (plans, reviews, commits) and hands work to a cheaper sidekick subagent through a plain-language brief. Cognition calls this Fusion.
Numbers (reported by the vendors, on their own benchmark and product)
Cognition, 3,000 sessions on FrontierCode 1.1:
| Config | Score | Cost per run | Lead turns | Lead input |
|---|---|---|---|---|
| Fable 5 (low) + sidekick | 60.7 | $1.86 | 11.5 | 545k |
| Opus 4.8 (medium) + sidekick | 54.6 | $2.04 | 26.5 | 1,679k |
| Fable 5 (low) alone | 60.8 | $4.03 | ||
| Opus 4.8 (medium) alone | 55.4 | $3.06 |
Fable costs twice as much per token and still came out cheaper. In 81% of Fable-led runs the lead never edited code (Opus: 24%), and in 13% it never read a repo file.
The finding is management style, not delegation count
Both leads handed off about 3 times per run. Fable delegated early, sometimes as its first action, with a brief that read like a design doc. Opus explored and built for 20 to 45 turns and handed off only the mechanical tail. On one task the spec required a hash that is O(1) in pointer length. Opus coded it itself, lost the constraint and shipped linear time, scoring 25. Fable’s brief said “operator() must be O(1) in pointer length: NO full token scan”, and the sidekick scored 94. After a handoff, Opus pulled the sidekick’s files back into its own context 2x as often and made 4x as many corrective edits at lead prices. Forcing Opus to delegate more made it worse: knowing what is safe to hand off is a judgment of its own.
What Factory adds: the cache
Providers cache the processed prefix per model, and reads from that cache cost about a tenth of fresh input. Moving the session to another model reprocesses the whole history at fresh rates. In Factory’s data, cache-blind switching reached 2.12x the all-frontier cost at turns 61-150 and 2.37x at turns 151-200; cache-aware routing stayed at 0.19x to 0.28x. Over two months of production, routing cut aggregate cost 58%, the median session 76%, and median turn latency from 81 s to 49 s. “Staying put is itself a routing decision.” A subagent with a short spec sidesteps the cache problem: the parent stays warm. Factory’s default reviewer comes from a different model family than the implementer.
Rule
- Delegate implementation, keep judgment. The brief lists constraints, edge cases and a definition of done; review the diff and re-delegate a fix instead of rewriting at lead prices.
- Delegate early or not at all. A late handoff of the tail saves little.
- Never switch the main session’s model mid-task to save money. Spawn a worker with a spec instead.
- Choose the lead by turns and context per task, not by price per token.
How it changes two pages here
cheap-model-delegation-boundary says editing cannot be delegated. That holds when the worker returns a summary, which carries no line numbers. Cognition’s sidekick reads and edits the files itself and the lead reviews a diff, and the score held. So the boundary is the summary, not the edit. Reasoning stays with the lead: serial debugging gained nothing, which matches Portal’s missed thread-safety bug.
commerce-agent-blueprint says to use no subagents, because every handoff is state-lossy. For a conversational flow, where the state is the product, that stands. For coding work with a delegable chunk, a brief that carries the constraints makes the loss cheap: use a sidekick there.
Exceptions
- Short tasks and serial debugging: nothing to hand off, so no saving.
- FrontierCode is Cognition’s own benchmark and Factory Router is Factory’s product. None of this has been reproduced independently.
- Lead behaviour is a property of a model release and can change with the next one.
Learned from
Joon Hee Lee, “Making Fable Cheaper Than Opus”, Cognition, 13 July 2026. Abhay Singhal, “Why model routing must be in the harness”, Factory, 24 August 2026. Related: claude-5-5-effort-and-cost (subagents moved to Sonnet 5.5), harness-cost-by-model-tier (making the cheap worker reliable), needle-tiny-tool-model.