System-1 gates System-2: the economics of routing, and buyers acting on it

Sources: arXiv:2610.39497 (Does Scaling Toward AGI Depend on What System-1 Lets System-2 Execute?), arXiv:2610.42053 (REA), arXiv:2610.39006 (Blind Spots, DAIR pick); OpenRouter, Arena and The Information for market evidence · raw: raw/huggingface/2026-11-02-does-scaling-toward-*.md, raw/huggingface/2026-11-02-rea-*.md, raw/dair/2026-11-02-230356-dair-academy.md, raw/rss/2026-11-02-the-information-it-s-contract-*.md

TL;DR

The System-1 paper formalizes routing as predict-gate-execute-update: a light predictor scores candidate interactions, an adaptive threshold admits some to an expensive executor, outcomes retrain the predictor. With ideal screening and equal costs, execution compute falls by up to 1/ρ_u, where ρ_u is the density of utility-relevant interactions; selection regret is bounded by predictor error. It proposes "informative compute" as the scaling coordinate. REA (embodied) applies the shape with route memory and a boundary-triggered judge: 95.9% on 3,000 LIBERO-Pro episodes vs 94.7% best single policy, 25% fewer control steps, 17% less control time. Blind Spots warns small models are weaker selectors: they recover 28-31% of the random-to-oracle gap vs 53% for GLM-5.2, and are more overconfident (Brier 29.1% worse).

Predict first, spend only on what passes the gate

The loop closes: outcomes retrain the cheap predictor.

flowchart LR
  C["Candidates<br/><small>calls, tools, steps</small>"] --> P["System-1<br/><small>cheap utility guess</small>"]
  P --> G{"Gate<br/><small>adaptive threshold</small>"}
  G -->|admit| E["System-2<br/><small>expensive executor</small>"]
  G -->|reject| K["Skipped<br/><small>compute saved</small>"]
  E --> O["Outcome<br/><small>measured utility</small>"]
  O -->|update| P
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  class C input
  class P,G loop
  class E core
  class K,O exit
  linkStyle 3 stroke:#2f9e44,stroke-width:2px
  linkStyle 5 stroke:#f08c00,stroke-width:2px

Blue is the candidate pool, amber is the cheap predictor and gate, purple is the expensive model, green is saved compute and feedback.

Market evidence (same day)

OpenRouter: users spending $100+/month use a median of 6 model families, double a year ago.
The Information: enterprises hesitate to raise annual commits because the best model changes week to week.
Arena Agent Arena: same intelligence can cost 10x more, need 6x more output tokens, take 2x longer; output tokens barely predict the bill.

Relation to prior wiki pages

Gives the theory behind LLM routing results: AgentBaton (10-30, capable-to-cheap handoff cut API cost 65.9%), small decision-model routers (10-28), Routed Continual Learning (11-01).
Tension with Blind Spots: a cheap gate is also a worse judge, so the ρ_u bound assumes predictor quality that small models may not have.
Connects to test-time compute allocation: the gate is an allocation policy over calls rather than tokens.