Assign models to roles by what each role’s decisions are worth. The most capable and most expensive model decomposes the work, resolves conflicts, chooses the path, and does the final review. Cheaper and faster models do the bounded, token-heavy work: repository scans, docs extraction, log reduction, test runs, browser passes, and mechanical edits. The expensive model re-verifies whatever it relies on from them.
builder-efficient-frontier puts it in one line: “Use the expensive frontier model where its marginal judgment matters. Push repeatable, bounded, or token-heavy work to cheaper/faster subagents.” builder-efficient-fable gives the same split for Claude Fable specifically and puts the payoff at “up to 3-5x more cost-efficient and 2-4x faster” when the work parallelizes. It labels that figure a workload-dependent estimate.
What keeps it from being plain delegation is that the trust boundary follows the price boundary. Cheap-model output is “leads, not facts”. “Before using a high-impact finding, opening a PR, or telling the user the work is done,” the expensive model reopens the cited files and checks the line refs. Without that step, tiering just moves the errors down to the cheaper model. The pattern depends on pattern-fresh-context-subagents (self-contained handoff packets, since the cheap agent “has no useful chat context”) and on pattern-evidence-before-claims (the re-verification step).
How it differs from pattern-cross-model-review. Cross-model review uses model diversity to find errors: different vendors look at the same artifact. Model tiering uses model price to allocate work: different tiers do different jobs. A framework can do both. pstack does, with a per-role model map and a cross-vendor review panel.
Applied by (backlinks)
builder-skills:
- builder-efficient-fable names Fable as orchestrator, architect, synthesizer, and final judge, with cheaper subagents for research, coding, testing, and log reduction.
- builder-efficient-frontier generalizes this to any high-cost frontier model, with the guardrails do not delegate the immediate blocker and do not let two agents edit the same files.
pstack:
- pstack-setup-pstack maps each role (architect, reviewer, worker, …) to a model and records the map in an always-applied rule. This is tiering done once as configuration rather than decided per task.
GSD:
- gsd-planner runs on an Opus-tier model, while execution goes to other agents. This is the earliest example in the wiki, implicit in which model each sub-agent runs on.
Provided by (harness)
- factory-droid has multi-model routing as a harness primitive. It routes subagent work to a model tier by task complexity, swaps models mid-session, and offers a free open-weight tier (Droid Core) plus BYOK. The harness does the tiering itself, so a framework does not have to build it out of subagent definitions.
See Also
- pattern-cross-model-review — the neighbour, which uses diversity for error-finding rather than price for allocation.
- pattern-fresh-context-subagents — the handoff discipline tiering depends on.
- pattern-context-engineering — keeping the expensive model’s context for decisions, not raw material.
- builder-stay-within-limits — the budget-window guard that pairs with it in builder-skills.