AI Benchmarks & Leaderboard — 2026-11-04

New Model Releases & Updates

Qwen 4 by Alibaba

• Type: Closed/Proprietary (Open-weights unconfirmed)
• Key benchmarks: No specific benchmark scores released yet; performance details pending.
• vs. Previous best: Successor to Qwen 3.8, which was a top open-weight model.
• What's notable: Four variants have been named, but Alibaba has not confirmed licenses or open weights, raising doubts about their "open-source promise" despite a history of 470+ open models.

  

Mistral Large 4 (Le Chonk) by Mistral

• Type: Open-weights
• Key benchmarks: High Artificial Analysis scores for coding.
• vs. Previous best: Ranks alongside Codex GPT-6.1 Terra in coding capabilities.
• What's notable: Weights have been released, making it a top choice for developers needing high-performance coding models that can be locally hosted or fine-tuned.

  

1 sources

zdnet.com

zdnet.com

Open-Weight Flagship Competition

• Models: Kimi K3, DeepSeek V4.1 Flash, Qwen3.8, GLM-5.3, and others are competing directly for developer budget.
• Notable: None is the "obvious pick," indicating a fragmented landscape where specific use cases determine the best model rather than a single dominant open-source leader.

  

Leaderboard Snapshot

Frontier Models (Closed-Source)

Note: Scores based on Artificial Analysis Intelligence Index.

Open-Source Leaders

Note: Specific aggregate scores for open models vary by benchmark; Kimi's drop is highlighted in recent news.

Benchmark Deep Dive

The most significant shift this week involves the Artificial Analysis leaderboard's methodology update, which has dealt a blow to Chinese contenders. Specifically, Kimi K3.1 Pro, previously ranked as the fourth-best AI model globally, has fallen out of the top tier. This change stems from Artificial Analysis revamping its tests to include private evaluations that are harder to "game" or overfit to public benchmarks.

This update raises critical questions about the true state of open-weight AI performance versus closed-source frontier models. The new data suggests that the perceived threat from Chinese labs to US giants like OpenAI and Anthropic may have been overblown when measured against these more rigorous, private tests. The gap between the top closed-source models (Gemini 4 Carbon, Claude Opus 5.5) and the open-weight leaders appears wider than previously thought in terms of raw intelligence index scores.

For practitioners, this means relying solely on public benchmark claims may be misleading. The shift toward private testing highlights the importance of independent verification and real-world application testing rather than just leaderboard chasing. It also signals that closed-source models are maintaining a significant edge in general reasoning and complex task execution.

Analysis & Trends

• State of the art: Google's Gemini 4 Carbon and Anthropic's Claude Opus 5.5 remain the dominant forces in closed-source intelligence, with GPT-6 Astra close behind.
• Open vs. Closed gap: The gap appears to be widening slightly in terms of verified "intelligence" due to stricter evaluation methods, though open models like Mistral Large 4 are closing the gap in specific niches like coding.
• Cost-performance: Inference costs for models comparable to early GPT-4 have plummeted, shifting the competitive advantage from raw model capability to application-level efficiency and cost management.
• Emerging patterns: There is a growing market for specialized "decision models" beyond traditional LLMs, offering alternatives for specific logical tasks.

What to Watch Next

• Qwen 4 Open Weights Release: Will Alibaba follow through with open weights for its 4 named variants, or will they remain proprietary?
• Artificial Analysis Index V5 Impact: Continued analysis of how the new private testing methods affect the rankings of other Chinese models like DeepSeek and GLM.
• Specialized Decision Models: Growth in the adoption of non-generative AI decision models as developers seek alternatives to traditional LLMs for specific workflows.