model-router

Model router — the right model for each agent

Multi-agent quality comes mostly from spending tokens across parallel contexts — so spend them where they matter and save them everywhere else. Set a model per subagent (the model field / the Task model option): opus · sonnet · haiku · fable · a full id · or inherit (default).

Role → model

Role / task character	Model	Why
Lead / orchestrator, synthesis, architecture, ambiguous hard reasoning	Opus	Control-logic quality drives the whole run
Implementation, refactors, writing tests, most production coding, most reviews	Sonnet	Best capability/cost balance — the production default
Search / discovery, grep-and-summarize, mechanical transforms, log triage	Haiku	~5× cheaper input than Opus; quality gap negligible on non-reasoning work
Security-critical or final whole-branch review	Opus	Worth the spend at the gate
Sub-task that needs your full context	inherit / fork	Keeps continuity; a fork reuses the prompt cache (cheaper)

The rule that decides the close calls

Turn count beats token price. A weak model that retries five times costs more — in tokens and wall-clock — than a stronger fast model that nails it once. When unsure between two tiers for real work, pick the stronger one; only drop to Haiku for genuinely mechanical tasks.

Two rules that stop silent overspend:

• Least powerful that suffices. Default a role down to the cheapest tier that can actually do it — the turn-count rule overrides only for genuine reasoning work. (One axis decides up, the other down.)
• Always set the model on a dispatch. An omitted model inherits the session default (often Opus) — so a grunt task silently runs on the priciest tier, defeating the whole point of routing. Name the model every time.
• Implementation tier by signal: 1–2 files + a complete spec → cheapest; multi-file / integration → mid; design judgment or broad-codebase reasoning → top. Never drop a reviewer (or an implementer working from prose, not a precise spec) below the mid tier.

Also dial effort

Independently of the model, set reasoning effort per subagent (low/medium/high/xhigh/max). Use low for cheap mechanical stages; reserve high+ for the hardest verify/judge/synthesis stages. Cheap model and low effort for the grunt work; strong model and high effort for the crux.

Cost anchors (per MTok, input/output)

Opus 4.8 $5 / $25 · Sonnet 4.6 $3 / $15 · Haiku 4.5 $1 / $5. So Haiku input is ~5× cheaper than Opus, Sonnet ~1.7× cheaper. A common, strong shape: Opus lead + Sonnet workers + Haiku scouts.

Defaults & cautions

• Don't override the model without a reason — inherit is right more often than not.
• The built-in Explore subagent already runs on a fast model and is read-only — prefer it for "where is X / how does Y work" instead of spinning up a custom researcher.
• Match effort and model together; a max-effort Haiku on a reasoning task is a false economy.

---

Credits / further reading: Anthropic pricing & subagents docs; the "model-per-role / turn-count-beats-price" framing is crisp in superpowers (obra/superpowers, MIT). See docs/ECOSYSTEM.md. Pair with subagent-orchestration.