{"slug": "gemini-4-argon-tops-new-hle-diamond-benchmark-on-closed-ended-research-questions", "title": "Gemini 4 Argon Tops New HLE-Diamond Benchmark On Closed-Ended Research Questions", "summary": "Google's Gemini 4 Argon leads Artificial Analysis's new HLE-Diamond leaderboard with 69.1%, ahead of Claude Opus 5.5 (max with fallback) at 67.9% and GPT-6 Astra (max) at 63.8%, on a 1,000-question closed-ended benchmark built by the Center for AI Safety and Scale AI from Humanity's Last Exam. Only 17 model variants, roughly 13% of those tested, score above 50%, and over 90% of models score noticeably higher on the reasoning partition than on knowledge, with Gemini 4 Argon at 79.6% reasoning versus 58.6% knowledge. The top open-weights model, Kimi K3.1 Pro at 37.3%, trails the leader by more than 30 points, and Claude Opus 5.5 gains 26 points moving from low effort (42%) to max effort (68%) at roughly 35 times the cost per task.", "body_md": "Artificial Analysis has launched HLE-Diamond, an independent leaderboard built on a refined, 1,000-question version of Humanity’s Last Exam, as part of the [v5 overhaul of its Intelligence Index](https://officechai.com/ai/artificial-analysis-launches-intelligence-index-v5-with-majority-private-datasets-openai-models-gain-muse-wonder-kimi-drop/). Google’s top model leads the new chart, with Anthropic close behind in second place.\n\nHLE-Diamond was built by the Center for AI Safety and Scale AI after a year spent cleaning up and refining the original Humanity’s Last Exam dataset with input from the research community. It comprises 1,000 closed-ended questions split evenly between reasoning and knowledge, 272 of which include an image. Artificial Analysis runs every question once without tool access, using the official HLE system prompt, and has GPT-6 Luna (medium) grade each response against the reference answer using the official HLE judge prompt, reporting the share of questions answered correctly (pass@1).\n\n## Gemini Leads, Claude Close Behind\n\nOnly 17 model variants — roughly 13% of those tested — score above 50% on HLE-Diamond. Google’s model sits at the top with a score of 69.1%, followed by Claude Opus 5.5 (max with fallback) at 67.9%. GPT-6 Astra (max) takes third at 63.8%, ahead of Claude Fable 5.1 (max with fallback) at 60.1% and GPT-6.1 Sol (max) at 58.8%.\n\nThe gap to the rest of the field is steep. Claude Sonnet 5.5 (max with fallback) rounds out the top six at 53.3%, while Meta’s Muse Wonder Full (max) and GPT-6.1 Terra (max) are the next-closest at 48.5% and 47.3% respectively — more than 20 points behind the leader.\n\n## Reasoning Beats Knowledge Almost Everywhere\n\nThe benchmark’s headline finding is how lopsided performance is between its two partitions. Over 90% of models tested score noticeably higher on reasoning questions than on knowledge questions. Google’s leading model scores 79.6% on reasoning against 58.6% on knowledge, while Claude Opus 5.5 and GPT-6 Astra are tied at 78% on the reasoning partition — though Opus 5.5 pulls ahead on knowledge, 57.8% to 49.6%.\n\nThe widest split among the top performers belongs to GPT-6.1 Sol (max), which scores 76.2% on reasoning but just 41.4% on knowledge — a 35-point gap. Across every model that clears 50% overall, the reasoning-to-knowledge gap runs between 16 and 35 points, underscoring that even frontier models still lean heavily on inference over recall when the questions get genuinely hard.\n\n## Open Weights Models Still Trail By Over 30 Points\n\nThe gap between closed and open models remains substantial on HLE-Diamond. [Kimi K3.1 Pro](https://officechai.com/ai/kimi-k3-1-pro-places-4th-on-artificial-analysis-intelligence-index-ahead-of-gpt-6-astra-gemini-4-argon/), whose weights have not yet been released, is the top open-weights model at 37.3%, followed by MiniMax-M3.1-Flash at 28.1% and Kimi K3 (max) at 27%. That leaves the best open model more than 30 points behind the leading proprietary systems.\n\n## Effort Settings Change The Score Dramatically\n\nHLE-Diamond scores climb sharply as models are pushed to higher reasoning-effort settings, but the cost climbs faster still. Claude Opus 5.5 gains 26 points moving from its low-effort setting (42%) to max effort (68%), at roughly 35 times the cost per task. GPT-6 Astra sees a comparable pattern, gaining 19 points from low effort (45%) to max (64%) at around 11 times the cost — a reminder that headline leaderboard numbers often come from the most expensive configuration a model can run, not the one most teams will deploy by default.\n\nThe release continues a run of fast-moving updates to the Intelligence Index family this year, following the [v4.2 rejig that had Claude Fable 5.1 on top](https://officechai.com/ai/artificial-analysis-rejigs-intelligence-index-with-v4-2-fable-5-1-tops-gpt-6-astra-placed-second/) and the subsequent run where [Claude Sonnet 5.5 landed three points ahead of GPT-6 Astra](https://officechai.com/ai/claude-sonnet-5-5-scores-56-on-artificial-analysis-intelligence-index-3-points-ahead-of-gpt-6-astra/) on the broader composite score.", "url": "https://wpnews.pro/news/gemini-4-argon-tops-new-hle-diamond-benchmark-on-closed-ended-research-questions", "canonical_source": "https://officechai.com/ai/gemini-4-argon-tops-new-hle-diamond-benchmark-on-closed-ended-research-questions/", "published_at": "2026-10-30 05:51:08+00:00", "updated_at": "2026-10-30 07:17:29.795316+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-research", "generative-ai"], "entities": ["Artificial Analysis", "Gemini 4 Argon", "Google", "Claude Opus 5.5", "Anthropic", "GPT-6 Astra", "OpenAI", "Kimi K3.1 Pro"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/gemini-4-argon-tops-new-hle-diamond-benchmark-on-closed-ended-research-questions", "markdown": "https://wpnews.pro/news/gemini-4-argon-tops-new-hle-diamond-benchmark-on-closed-ended-research-questions.md", "text": "https://wpnews.pro/news/gemini-4-argon-tops-new-hle-diamond-benchmark-on-closed-ended-research-questions.txt", "jsonld": "https://wpnews.pro/news/gemini-4-argon-tops-new-hle-diamond-benchmark-on-closed-ended-research-questions.jsonld"}}