Gemini 4 Carbon tops Artificial Analysis' new HLE-Diamond at 69.1%, and places fourth on Cognition's FrontierCode 2.0

Two independent boards scored Gemini 4 Carbon this week, and they point different ways. Artificial Analysis launched HLE-Diamond on October 30, its own run of a refined 1,000-question version of Humanity's Last Exam, and Carbon (high) leads at 69.1%, 1.2 points ahead of Claude Opus 5.5 (max) at 67.9%. On Cognition's FrontierCode 2.0 coding board, Carbon places fourth at 40.0%, behind Claude Opus 5.5, GPT-6 Astra and Claude Sonnet 5.5. We read both leaderboards directly.

HLE-DIAMOND

- Overall: Gemini 4 Carbon (high) 69.1%, Claude Opus 5.5 (max) 67.9%, Claude Opus 5.5 (xhigh) 64.0%. - Reasoning half: Carbon 79.6%, with Opus 5.5 (max) and GPT-6 Astra (max) both at 78.0%. - Knowledge half: Carbon 58.6%, Opus 5.5 (max) 57.8%. - Artificial Analysis' announcement says only 17 model variants, about 13% of those tested, score above 50%. The best open-weights model is Kimi K3.1 Pro at 37%, more than 30 points back, and its weights are not out yet.

HOW IT IS RUN

HLE-Diamond was released by the Center for AI Safety and Scale AI after a year of cleaning up the original Humanity's Last Exam. It has 1,000 closed-ended questions, split evenly between reasoning and knowledge, and 272 of them include an image. Artificial Analysis runs every question once with no tools and the official HLE system prompt, and GPT-5.6 Luna grades each answer with the official judge prompt. It is part of v5 of the Artificial Analysis Intelligence Index.

FRONTIERCODE 2.0

Cognition's benchmark asks whether a repository's maintainer would actually merge the agent's change, across 132 tasks written by open-source maintainers. Cognition added Carbon on October 28.

- Claude Opus 5.5 (max, Claude Code): 45.5%, $12.82 a task. - GPT-6 Astra (max, Codex): 42.8%, $6.68. - Claude Sonnet 5.5 (max, Claude Code): 41.0%, $19.94. - Gemini 4 Carbon (medium, Devin CLI): 40.0%, $4.04. - Gemini 4 Argon (high, Devin CLI): 26.3%, $6.51, twelfth of 26.

Cost is Cognition's mean spend per run at list API prices. Each lab's models run in a different agent harness, so a row measures the model and its harness together.

WHAT WE DO NOT KNOW

When anyone outside Google's trusted programs can use Carbon. Access is still limited to cyber defenders and Fairwind partners, Artificial Analysis marks it as having no publicly available API, and Google has given no date for developers. Cognition's post on X gives Argon's cost as $6.41 a task; its leaderboard says $6.51, and we use the leaderboard.