AAI Index Analysis: Luna Leads Cost-Efficiency, Qwen 3.8 27B Ranks Fourth

On August 20, MikePFrank posted a series of threads ranking models by "price-performance" (Intelligence Index score divided by per-task cost), based on public data from the Artificial Analysis (AAI) Index, revising his calculation method along the way.

Confirmed

• Among models scoring at or above Qwen 3.8 27B, Luna is the most cost-efficient model per unit cost, followed closely by DeepSeek V4 Flash, with the newly released Qwen-3.8 27B ranking fourth on this metric.
• Qwen-3.8 27B's unique advantage is that it is the only model on the list that can run locally on a laptop — a true one-of-a-kind.
• The author corrected the "per-cost Agent Index" calculation column, noting it should be recalculated based on the per-run test cost of that index; the updated data table only affects the ranking of a few models.
• On the agent cost-efficiency Pareto frontier chart drawn from AA data, Qwen-3.8 27B sits on the frontier, while Claude does not.
• AAI provides Pareto frontier charts of Intelligence Index scores versus cost.

Why it matters

• The author argues that Pareto frontier charts show the performance-versus-cost trade-off more intuitively than plain rankings, making them more useful for developers choosing models under different budgets.
• The analysis brings "local deployability" into the price-performance discussion, highlighting what small-scale models like Qwen-3.8 27B mean for local developers.

2026-08-20 ~ 2026-08-20 · 5 related posts

• [source] Luna leads cost-efficiency, Qwen 27B stands out as laptop-viable option — MikePFrank · 2026-08-20
• Luna Remains Most Cost-Efficient, Qwen 3.8 27B Ranks Fourth — MikePFrank · 2026-08-20
• AAI Publishes Pareto Frontier for Intelligence Index vs Cost — MikePFrank · 2026-08-20
• Correcting Agentic Index: Introducing Cost-Per-Test and Pareto Frontiers Analysis — MikePFrank · 2026-08-20
• [source] Agent cost-efficiency plot: Qwen hits Pareto frontier — MikePFrank · 2026-08-20