September 2026 AI LLM Rankings: Claude and OpenAI Battle for the Top Spot

Overview

In mid-September 2026, the field of large AI models welcomed a new round of technological iteration and reshaping of the landscape. With Anthropic releasing the Claude Fable 5.1 series and OpenAI launching GPT-6 Astra, the competition among top-tier models has intensified. Although there is currently no universally recognized "comprehensive ranking" standard in the industry, the subtle differences between the capability boundaries of current mainstream models and users' real-world experiences can be clearly observed through the dual perspectives of Artificial Analysis's Intelligence Index and LMArena (formerly Chatbot Arena)'s Elo ratings.

"Objective benchmark tests measure a model's theoretical ceiling, while crowdsourced blind tests reflect users' genuine preferences in complex scenarios."

Dual-Track Analysis of Comprehensive Rankings

In the Artificial Analysis Intelligence Index (AA Intelligence Index), Anthropic's Claude Fable 5.1 leads with a perfect score of 53, closely followed by OpenAI's GPT-6 Astra (max), which also scores 53. The two are neck-and-neck across 10 standardized evaluations including programming, mathematics, and reasoning. Notably, the Claude Opus 5 series also occupies a prominent position on the leaderboard, demonstrating Anthropic's deep accumulation in high-end reasoning.

In contrast, the LMArena Elo rating based on anonymous A/B voting presents a different map of user preferences. Claude Fable 5 ranks first with an Elo score of 1506, while its predecessor, the Claude Opus 4.6 series, also performs strongly. Of particular note is that Facebook AI Research Lab's Muse Spark 1.2 (xhigh) enters the top five for the first time, surpassing flagship models from multiple companies with a score of 1499, marking a breakthrough in the open-source camp at the level of user experience.

Sub-fields and the Rise of Domestic Models

In niche tracks, DeepSeek V4.1 Flash emerges as a dark horse on the coding ability leaderboard, ranking first and demonstrating the significant advantages of domestic large models in programming assistance. Meanwhile, Gemini 3 Deep Think and DeepSeek V3.2 Speciale follow closely behind, making competition in the coding ecosystem increasingly fierce.

In terms of mathematics and Agent capabilities, xAI's Grok 4 Heavy tops the mathematics ranking, while OpenAI's o3-pro narrowly leads Moonshot AI's Kimi K3 to take second place on the Agent leaderboard. Data shows that Alibaba's Qwen3.8-27B newly enters the Agent ranking at fourth place, while Zhipu AI's GLM-5.3 also secures sixth place in comprehensive evaluations. The performance of leading domestic models in specific tasks has already formed direct competitive dynamics with the international first-tier echelon.

Outlook on Market Landscape

The current SOTA (State of the Art) title frequently rotates between Claude Fable 5.1 and DeepSeek-V4.1-Flash. The former holds the crown of comprehensive performance with a high HLE score of 65.00, while the latter leads the open-source and domestic model camps with a score of 63.90. With September data updates, the fourth quarter is expected to see the release of more next-generation models optimized based on reinforcement learning and long chain-of-thought (Long-CoT). The comprehensive ranking of global AI large models may undergo drastic fluctuations again.