When Google released Gemini 4 Argon, its announcement included scores like 77.9% on the DeepSWE coding benchmark — but those numbers were vendor-reported, meaning Google ran the tests itself.

Benchmarks are standardized tests — coding challenges, reasoning puzzles, professional exams — that let labs compare models on a common scale. The problem is that a lab can choose which benchmarks to highlight, tune specifically for the ones it knows are coming, or run tests in favorable configurations. Independent evaluators often get different numbers.

This is why third-party leaderboards matter: organizations like Artificial Analysis run every major model through the same tests under the same conditions. Gemini 4 Argon scored 52.6 on the Artificial Analysis Intelligence Index — a notably different picture from its vendor-reported figures. Neither number is a lie; they just measure different things. The gap between them is where the marketing lives.