Model-Tier Routing

Within a single task, split the work across model tiers by role: put the expensive/smart model on the high-leverage cognitive role (architecture, planning, deciding which model does what) and delegate bulk execution, breadth, and research to cheaper models. The lever is that a premium model billed by usage is wasteful when made to do everything — plan and execute and gather context — because most of that work is low-intellect labor a cheaper model does about as well for a fraction of the cost. The discipline is to route each sub-role to the cheapest tier that still clears the bar, and reserve the top tier for the small slice of work that genuinely needs it. This is Reasoning Effort Control generalized from one dial on one model to which model runs which part.

The source ("Chase AI", on reducing Fable 5 usage) shows the pattern in three directions:

Smart model as architect, cheap models execute (dispatch down). Fable 5 writes the plan, then — by complexity — hands each part to the appropriate model: Opus, Sonnet, an outside model (GPT-5.5), or something local. Fable can name the executor per step in the plan itself. The low-ceremony version: plan mode → a markdown plan → spin up a separate Opus session to execute it, so Fable doesn't burn tokens on low-level implementation.
Cheap models research, smart model plans (gather up). Inverse direction: because a premium model's knowledge cutoff isn't current and research "doesn't require super-high reasoning," let cheaper models (Opus/Sonnet sub-agents in a deep-research fan-out) gather web context and do baseline adversarial vetting, then hand that context to the smart model to plan from. Running the premium model as every sub-agent in a 100+-agent fan-out would blow through usage limits for no benefit.
Cross-vendor delegation as a routing target. The cheap tier can be another vendor's model reached through a bridge (the source names the Codex plugin for Claude Code: /codex:rescue, /codex:transfer to delegate work to GPT-5.5). Routing is model- and vendor-agnostic.

The productized, escalation-based form of this split — a cheap executor that phones a smart advisor only when stuck — is Advisor Mode.

Claims

Assign model tiers by role within a task: the expensive model does the high-leverage cognitive work (architecture, planning, deciding the routing); cheaper models do bulk execution and research. principle — durable: the cost/capability tradeoff that governs a single model's effort dial also governs which model runs each sub-role, so spending the top tier on low-intellect labor is the same overkill as maxing effort on a trivial task. This is Reasoning Effort Control lifted to the fleet.
Have the planner name the executor per step — "Opus for part one, Sonnet for part two, send part three to GPT-5.5" — rather than running one model for everything. best practice — context: multi-step agentic work where step complexity varies; the model is "smart enough to know which model is best for the job," but the routing only pays off when steps genuinely differ in required capability.
The low-ceremony version needs no orchestration harness: plan mode → a markdown plan → a fresh session on a cheaper model executes it. best practice — context: operators who don't want to wire sub-agent dispatch; a hand-off via a written plan captures most of the savings ("you don't need to overcomplicate it"). Trades some automation for simplicity — see Agentic Simplicity.
Don't run the premium model as every sub-agent in a large fan-out — route breadth (research, context-gathering, adversarial vetting) to cheaper tiers and reserve the premium model for the planning/architecture step. best practice — context: dynamic-workflow fan-outs (the source's deep-research run spawned 109 sub-agents); using the top model for all of them "makes no sense" and exhausts usage limits, but the direction (cheap-gathers → smart-plans) assumes the research tier is capable enough to vet its own findings. See Workflows vs Agents, Agentic Workflow Patterns.
Routing targets can be cross-vendor via a bridge (the source names a Codex plugin: /codex:rescue, /codex:transfer to delegate to GPT-5.5). observation — a named mechanism, groundable; the point is that "cheaper tier" isn't limited to one vendor's ladder.
Split ingest from synthesis by tier: use a cheaper-but-adequate model for the mechanical bulk ingest, and reserve the premium model for the hard synthesis step. best practice — context: building a Compiled Knowledge Base, where reading a source into wiki pages is high-volume low-intellect labor but the audience-shaping/creative pass is where capability pays. A second, independent source (Nate Herk) says the premium model "is probably overkill for the ingest — switch that to Opus," then reserves it for the framing step; corroborates "route each sub-role to the cheapest tier that clears the bar." The paired model-comparison claim (that the premium model understood an "emotional / beginner-friendly" framing better than another top model) is the source's assertion — see the distillate's Key claims, groundable, not adjudicated here.
The bottom tier can be self-hosted: open-weight models running locally or on an aggregator are a routing target, reachable because the harness is model-indifferent. best practice — context: an operator with hardware or a metered third-party endpoint. A third, independent source (Nate Herk, on Ollama + Claude Code) scopes this tier to the same task shape this page already names — "reading and summarizing files before passing to a smarter model," grepping a codebase for relevant functions, repetitive code scaffolding, first-pass triage of issues and error logs — plus two availability cases (the vendor is down; you hit a session limit). Three sources now agree on route each sub-role to the cheapest tier that clears the bar. See Harness / Model Separation, BYO Inference Endpoint.
The cheapest tier has a floor: below it, dollar savings are paid back in latency, lost tool-call visibility, and operator attention. principle — durable, and the necessary caveat to this page's central discipline. The bar a tier must clear is priced in attention, not only in dollars per token: the same source's 9B local model was nearly free and spent four minutes producing an unsupervised answer with no visible tool calls. See Total Cost of Inference.
A harness's internal sub-model slots (tool-call model, small-fast model, subagent model) are themselves a tier-routing surface — and one the vendor has already routed for you. observation — redirecting them is tier routing performed on someone else's map; leaving one unset silently bills the vendor's paid default. See BYO Inference Endpoint.
The productized form is a model router with an intent classifier: a fast, cheap classifier predicts what the user is trying to do and dispatches the query to the appropriate model or pipeline; because routers are cheap they can be stacked without adding significant latency or cost. observation — the AI Engineering textbook names routing as a distinct stage of architectural maturity, arriving at the same "cheap router, right model per request" pattern from the product-architecture angle; the companion model gateway (unified interface, fallback, load-balancing, access control) is held as a related-but-distinct pattern in the distillate.
Frame the split as an org chart: the expensive model is the boss/foreman — it writes the specs, designs the system, reviews the work, and rules on disputes, and it never codes — while the coding goes to the cheapest worker that has clear specs. best practice — context: multi-agent builds with checkable outputs. A fourth, independent source (Nate B Jones, rebuilding a real author website with a boss-plus-four-cheap-worker swarm) corroborates "reserve the top tier for planning/review, route execution down," and reports the same job at ~$8 all-in through the org chart vs an estimated $85–105 all-premium on 11–13M tokens — a 10×+ gap "and nothing got worse… Fable did more judging, not less." The source's framing: an over-budget AI project is usually an org-design failure (no router, the expensive model doing everything), not an AI failure. The staffing-time complement — qualifying an unfamiliar worker model for its role before routing to it — is Agent Audition. Pricing figures are the source's claims; see the distillate's Key claims.
The second slot need not be a lower tier: a cross-vendor peer can be selected for independence and for which meter it bills, not for being cheaper-and-dumber. best practice — context: an operator holding subscriptions with two vendors, on load-bearing work where a check's independence is worth paying for. A fifth source (Chase AI, grill-me-codex) routes the execution slot to OpenAI Codex while the premium model plans and reviews — but argues the rival model is stronger, not weaker, inverting this page's usual "cheapest that clears the bar" criterion. Two selection criteria are in play and they are not the same question: cost asks for the cheapest adequate model, independence asks for the least correlated one (Cross-Model Independence). They coincide in this source, which hides the distinction; when they diverge, trade them off explicitly.
The routing question has a prior: route only after deciding the task is worth agents at all. best practice — context: an operator reaching for a fleet. The same source (Nate B Jones, in a later video) restates the org chart — "an expensive model to plan and judge like Fable and a lot of cheap models to do all the coding and burn all the tokens," with the same ~10× reduction claim — but frames tier routing as what makes multi-agent affordable, not as what makes it right. The affordability is what raises the stakes on the prior question: cheap execution means the wasteful failure is no longer an obvious bill, it is a fleet quietly pointed at work that wanted a chat window or no AI at all. See Agent-Shape Triage, Team-Forming Constraints. Restatement by the same source as the org-chart claim above, so not independent corroboration; the 10× figure is check-worthy.
A cheaper tier can sometimes clear the bar by retrying rather than by being upgraded. best practice — context: tasks with a mechanical checker, where retries are cheap and gradeable. The same source reports a study in which a cheap model at 250 attempts beat the best single attempt of the best model available — which gives "cheapest tier that clears the bar" a second lever, since the bar is a function of attempts as well as capability. Sharply contingent on the checker: without one, extra attempts buy answers nobody can identify and the upgrade was the right call after all. See Repeated-Sampling Scaling. Figures are the source's claims; check-worthy.
What is scarce may be an allowance, not a dollar: routing across meters (separate vendor subscriptions) preserves a capped premium budget even when total spend doesn't fall. observation — the source's Claude Code banner reads "up to 50% of your weekly usage limit on Fable 5… Fable 5 draws down usage faster than Opus 4.8," and the payoff quoted is "~130,000 tokens on the Fable side" for a whole build. Offloading execution to another vendor's subscription moves burn off the constrained meter; whether it reduces cost is a separate, unmeasured claim. Held as a distinct idea in the distillate.

Reasoning Effort Control — the same cost/capability tradeoff on one model's effort dial; tier routing is that discipline generalized to which model runs each role.
Cross-Model Independence — the same seam with a different selection criterion: route to the least correlated model rather than the cheapest adequate one. Tier routing optimizes cost; independence optimizes what a checker can see.
Advisor Mode — the escalation-based instance: a cheap executor consults a smart advisor only when stuck, instead of the smart model dispatching work up front.
Layered Agentic Architecture — sub-agents let you specialize and scale; tier routing decides which model tier each specialized sub-agent should carry for its role.
Managed Agent — a demo CMA defaulting to the premium model and being switched down to a cheaper one is tier routing applied to a hosted loop's cost.
Safety Routing Fallback — a different routing axis: a classifier silently swaps in a cheaper model for the same request; tier routing deliberately splits different sub-roles across tiers.
Workflows vs Agents — deep-research fan-out is the workflow substrate where "cheap models gather, smart model plans" is enacted.
Agentic Simplicity — "route to the cheapest tier that clears the bar" is pay-for-capability-only- where-it-pays-off applied across models.
Compiled Knowledge Base — ingest (cheap tier) vs synthesis (premium tier) is tier routing applied to knowledge compilation.
Execution Commoditization — the economic why behind this routing: cheap execution is commoditizing, so the routing layer, alone, is table stakes; the durable value moves to the frontier question of what to build (see Imagination Constraint).
Harness / Model Separation — the architectural precondition: routing a sub-role to a cheaper tier requires the loop around it to be indifferent to which model answers.
BYO Inference Endpoint — the config seam that makes a self-hosted or third-party bottom tier real, and the partial-override trap that silently un-routes you.
Total Cost of Inference — the floor beneath the cheapest tier: attention is the currency the bar is really priced in.
Open-Weight Capability Gap — the gap is task-shaped, which is what makes an open bottom tier coherent rather than a uniform quality cut.
Agent Audition — the staffing-time gate: routing picks the tier for a role; the audition qualifies a specific unfamiliar model for that role before work is routed to it.
Check Gaming — the risk that comes with routing execution to the cheapest worker: "cheapest adequate" is also the worker most likely to game a weak check.
Agent-Shape Triage — the prior question this page assumes answered: routing decides which tier runs each role inside a shape already chosen. Cheap execution makes getting the shape wrong quieter, not cheaper.
Repeated-Sampling Scaling — the second lever on "cheapest tier that clears the bar": where a mechanical checker exists, retries can substitute for capability.
Distillate: Make Fable 5 80% Cheaper (& Other Usage Cheat Codes) — five levers for spending a premium model less
Distillate: Fable 5 + Karpathy's LLM Wiki is Basically Cheating — "premium model is overkill for ingest; switch ingest to a cheaper tier, keep it for the synthesis."
Distillate: You Can't Compete on Cheap Models Anymore — an economics-framed source that corroborates "route execution to the cheapest adequate tier" while arguing that layer is only table stakes: the leverage is the frontier "what's now possible?" question on top of it.
Distillate: Ollama + Claude Code = 99% CHEAPER — adds a self-hostable open-weight bottom tier, the same task-scoping from a third source, and the floor beneath "cheapest."
Distillate: AI Engineering in 76 Minutes — Chip Huyen's Book, Speedrun — textbook corroboration of the intent-classifier model router as a maturity stage, plus the companion model-gateway pattern.
Distillate: STORM: A Fixed Panel of Adversarial Research Lenses, Packaged as a Skill — lens subagents run on Opus 4.8 with Haiku/Sonnet named as cheaper options: per-subagent tier routing inside a fixed roster (and see Bounded Fan-Out for the count axis it corroborates on the 100+-agent fan-out).
Distillate: Claude Fable 5 Bossed 20 Cheap AI Agents. The Whole Site Cost $8. — the org-chart framing (expensive boss that never codes over cheap workers), a 10×+ cost gap on identical work, and "an over-budget AI project is an org-design problem, not an AI problem."
Distillate: Fable 5 + GPT 5.6 Sol = CHEAT CODE — routes the execution slot cross-vendor to a peer the source claims is stronger, not cheaper-and-dumber; surfaces meter-vs-dollar as separate scarcities, and breaks with Advisor Mode over which model fills the executor slot.
Distillate: 1.6M agents registered for OpenClaw and did NOTHING. — the same source restating the plan-and-judge-expensive / execute-cheap split (~10×) inside a named harness, plus the prior question routing doesn't answer (is this task agent-shaped at all?) and retry-instead-of-upgrade as a second way to clear the bar.

Model-Tier Routing

Claims

Related

Linked from