Artificial Analysis launches HLE-Diamond leaderboard, and Gemini 4 Carbon takes the top spot
Google's newest model leads a tougher 1,000-question AI benchmark, while open-weight models trail the leaders by more than 30 points
Artificial Analysis launched its HLE-Diamond leaderboard on October 27, 2026, and the early results show a clear hierarchy: Google’s Gemini 4 Carbon sits at the top with 69.1% overall accuracy, and open-weight models are a long way behind.
What HLE-Diamond actually measures
HLE-Diamond is a refined subset of Humanity’s Last Exam. The new version trims the test down to 1,000 questions.
Those questions are split between two categories: reasoning and knowledge. Reasoning tests whether a model can work through a problem. Knowledge tests whether it actually knows the relevant facts.
The set also includes 272 image-based items. That means a model cannot coast on text skills alone. It has to interpret visual information too.
Artificial Analysis is using HLE-Diamond to replace its earlier Humanity’s Last Exam evaluation. The stated goal is a more accurate picture of how AI models perform across realistic workloads.
The leaderboard: Google first, Anthropic close behind
Gemini 4 Carbon leads with 69.1% overall. Its strongest area is reasoning, where it scored 79.6%.
Knowledge was a different story. The model scored 58.6% on that half of the test, a gap of 21 points between its two categories.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Anthropic’s Claude Opus 5.5 took second place at 67.9%. That leaves just 1.2 points between the top two.
Below the leaders, the field thins out quickly. Only around 13% of the model variants tested cleared 50% accuracy on HLE-Diamond.
The open-weight problem
On HLE-Diamond, the gap between the top proprietary models and the best open-weight models exceeds 30 points.
That is not a rounding error. On a test where the leader scores 69.1%, a 30-plus point deficit puts the best open alternatives somewhere below 40%.
How it fits into the bigger index
HLE-Diamond is not a standalone project. It arrived as part of the rollout of the Artificial Analysis Intelligence Index v5.
Within that index, HLE-Diamond carries a 10% weight. It is one ingredient in a broader composite score, not the whole recipe.
Gemini 4 Carbon was announced on October 28, 2026, one day after the leaderboard went live. The model launched with introductory pricing of $2/$10 per million input/output tokens.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.

