September 2026 AI Large Model Rankings: Claude Fable 5.1 Tops the Charts as Domestic Models Accelerate Breakthroughs

September 2026 saw a new round of reshuffling in the global competitive landscape of large AI models. Drawing on multiple authoritative evaluation data sets, Anthropic’s Claude Fable 5.1 surged to the top, not only claiming the top three spots on the Artificial Analysis intelligence index but also winning user preference in LMArena blind tests. Meanwhile, domestic models represented by DeepSeek and Zhipu AI continued to make strides in the code, mathematics, and agent (Agent) tracks, demonstrating strong innovative vitality and momentum.

Overall Rankings: Anthropic Dominates the Charts, Closed-Source Models Lead the Top Tier

There is currently no single universally recognized “authoritative” ranking; evaluations typically combine objective benchmark tests with subjective user voting. According to the Artificial Analysis intelligence index, updated on September 13, 2026, Claude Fable 5.1 occupies the top spot with a high score of 53 points, closely followed by GPT-6 Astra (max), which also scored 53 points. The two models performed similarly across 10 standardized evaluations covering mathematics, programming, and reasoning.

In the LMArena (Chatbot Arena) rankings, which reflect real-world user experience, Claude Fable 5 leads with an Elo score of 1506 points, while Claude Opus 4.6 (high) takes second place, trailing by just one point. The data shows that Anthropic’s Fable series has established a significant advantage in balancing efficiency and quality, consistently remaining at the top of user preference.

“Objective benchmarks measure capability ceilings, while crowdsourced blind tests reflect real-world experience.” Industry insiders pointed out that the Claude series maintains a leading position under dual validation from both metrics, consolidating its status as the current strongest general-purpose model.

Breakthroughs by Domestic Models: DeepSeek V4.1 Flash Shines in Code

Although the top spots on the leaderboards are dominated by American tech giants, domestic models have achieved breakthrough progress in vertical fields. DeepSeek-V4.1 Flash emerged as the standout highlight of the month, ranking first on the code capabilities leaderboard upon its debut (scoring 3471) and earning dual titles of “Best Open-Source” and “Best Domestic Model.” Its HLE individual score reached 63.90, trailing only Claude Fable 5.1.

Additionally, Zhipu AI’s GLM-5.3 and Moonshot AI’s Kimi K3 delivered impressive performances in multimodal and Agent tasks. Kimi K3 ranked third on the Agent capabilities leaderboard (scoring 84.8), surpassing OpenAI’s o3-pro; meanwhile, Qwen3.8-27B, released by Alibaba, also entered the top five of the Agent leaderboard for the first time, marking that domestic models have joined the world’s vanguard in complex task collaboration capabilities.

Sub-Specialties: Differentiated Competition Between Open-Source and Closed-Source Models

In terms of parameter scale and licensing type, the market shows a polarized trend. Closed-source models such as Claude Fable 5.1 and GPT-6 Astra continue to monopolize the high-score segments of general intelligence evaluations like ARC-AGI-2. However, in vertical fields such as code and reasoning, open-source models are rapidly closing the gap. DeepSeek-V4.1-Flash offers a “free commercial use” license, significantly lowering the threshold for developers to access high-end AI capabilities, and is viewed as an important milestone for the open-source community.

In the area of mathematical capabilities, xAI’s Grok 4 Heavy leads with a perfect score of 100, while Step 3.5 Flash and Kimi K2 Thinking also entered the top ten, demonstrating the sustained investment by domestic teams in long-chain mathematical reasoning. As evaluation benchmarks evolve toward more practical metrics such as SWE-bench Pro and Terminal Bench 2.0, future rankings may be further reconstructed, providing greater room for advancement for models with genuine engineering deployment capabilities.