Artificial Analysis Updates AI Leaderboard Evals to Rebalance Scores

Artificial Analysis, an independent testing group, has released a major update to its primary AI performance benchmark. The new framework, known as Intelligence Index v5, relies heavily on secret test questions and multi-step work environments such as writing software and conducting research. By using hidden datasets rather than public ones, the benchmark aims to prevent tech companies from designing models that simply memorize exam answers.

According to the updated scores, the performance gap between top proprietary models and open-weights alternatives has widened dramatically. Leading systems from Anthropic and OpenAI held the top positions on practical tasks. For businesses and everyday users, this means that paying for subscription cloud services still delivers superior real-world capabilities, even though running costs among top-tier models vary by more than thirteen times.

What remains to be seen is how long these private test sets can stay secure. As benchmark creators rely more on hidden questions to stop software from gaming the system, consumers are forced to trust evaluation firms without being able to inspect the test material directly. It is also unclear whether open-source developers can catch up to these proprietary leaders without matching the massive hardware investments of major tech labs.