What does switching from a frontier model to a mix actually cost?
Prices updated: Aug 3, 2026
Based on official API token prices from the providers (as of Aug 3, 2026, see footnote). The calculator applies one fixed token profile per task across all models to keep them comparable — this illustrates the order of magnitude, it is not a calculator for real contracts.
Cost in selected mix
–
Cost at 100% expensive model
–
Savings
–
Calculation: Cost/task = (1,000 input tokens × input price + 3,000 output tokens × output price) ÷ 1,000,000 — an assumed token ratio typical of agentic workflows, applied identically to every model. Total cost = volume × (expensive share × cost/task expensive + affordable share × cost/task affordable). Token price sources: OpenAI (Jul 30, 2026), Anthropic, Google, DeepSeek, Zhipu, Moonshot — as of Aug 3, 2026, see research dossier.
As of Jul 30, 2026: OpenAI cut API prices for GPT-5.6 Luna by 80% and for Terra by 20%; Sol stayed unchanged. OpenAI says Luna beats Anthropic's Fable 5 on the "Agents' Last Exam" benchmark at roughly 99% lower cost per task. Anthropic had moved first a few days earlier (Jul 24) with Opus 5 — priced at half of Fable 5. (Source: OpenAI, Bloomberg)
Real-world example from the reporting: Cursor built a complete web browser once using only an expensive frontier model (cost: $10,000) and once using a model mix (cost: $1,339) — an 87% saving with comparable results. (WSJ, Jul 24, 2026)