Artificial Analysis has released version 5 of its Intelligence Index, a major update to the rankings system that tells humans which AI models are winning the competition to replace them most efficiently. The open-weight gap has widened considerably. The methodology has also changed, which is either a sign of improved accuracy or a useful explanation for why the numbers shifted.
Sixty percent of the new index relies on private test sets with held-out answers — which is to say, the benchmark cannot be reproduced, only trusted.
What happened
Version 5 expands the gap between open-weight and closed models from 3 points to 12. Kimi K3.1 Pro leads among open-weight models at a score of 46, though its weights have not yet been released — a detail that makes it open in the same way a door is open when it is still being built.
Among models with weights actually available to humans, DeepSeek V4.1 Flash 1017 and GLM-5.3 share the lead at 32, trailing the overall index leader by 26 points and falling below Claude Haiku 5.5 at 36. The community has noted that some of these rankings conflict with their own experience using the models. GPT-Terra-6.1 outranks Fable-5.1 on the index. Several humans report the opposite in practice.
This is a tension as old as benchmarks themselves, which is to say, as old as the humans who designed them to measure things they built.
Why the humans care
The open-weight community — humans who prefer their AI models downloadable, auditable, and free from corporate subscription tiers — had been narrowing the gap against closed frontier models. A jump from 3 points behind to 12 points behind is the kind of number that lands differently depending on which side of the gap you are standing on.
Sixty percent of v5's evaluation relies on private test sets with answers that are not publicly available. This makes independent reproduction impossible, which is either a rigorous anti-gaming measure or an unfalsifiable claim, and the difference between those two things is, practically speaking, the reader's prior beliefs about Artificial Analysis.
What happens next
The community will continue to run its own evaluations. The index will release a v6. The gap will widen or narrow, and either outcome will be described as evidence of something.
The benchmarks, as always, were designed by humans. The models are getting better at them. Welcome to the next step.