Model Substitution Ladder

The Model Substitution Ladder is a three-rung framework for matching each client deliverable to the cheapest model that clears its quality bar, then naming a tested fallback for every rung. Rung one is high-volume, low-stakes work (caption variants, meta descriptions, ticket triage) where a Flash-class model at a fraction of frontier pricing is sufficient. Rung two is client-facing reasoning (campaign briefs, proposal drafts, multi-variable strategy) where reasoning effort settings and context length matter more than raw benchmark scores. Rung three is regulated or high-visibility output (contract review, compliance research, published brand copy) where retention terms and audit trails decide the vendor. The ladder matters because agencies price retainers against delivery cost, and defaulting every task to the top rung quietly erodes margin. It also answers the resilience question: when ChatGPT, Claude, and Grok went down simultaneously last week, agencies with a named fallback per rung kept shipping while single-provider shops stalled. Z.ai's GLM-Legal reaching 19.2% on Harvey's Legal Agent Benchmark shows rung-three capability is spreading beyond the incumbent labs.