Artificial Analysis

Comparisons across the AI stack

Artificial Analysis benchmarks models and agents as well as the services and hardware used to run them. It describes itself as an independent benchmarking company, with founders Micah Hill-Smith and George Cameron. The organization supplies comparisons for developers and buyers choosing between systems, rather than producing the underlying frontier models. [1]

Communication style as a separate measurement

Its October 2026 AA-PersonalityBench release measures differences in default model responses. Models answer 900 questions without a system prompt; a panel of model judges compares responses on thirty traits, while twenty-one communication characteristics are measured programmatically. [2]

Reading a benchmark in context

The personality scores describe how strongly a trait appears, not whether that trait is desirable for every user. They are separate from general intelligence or factual correctness. A model can rank highly on one communication preference while being a poor choice for another task. [2]

The wider benchmarking role also spans infrastructure and operating tradeoffs. Readers comparing systems should retain the particular model, agent harness, serving setup and measurement being reported instead of treating every leaderboard position as the same kind of evidence. [1]