Google and Meta surpass Anthropic in AI rankings

In October 2026, Google and Meta surpassed Anthropic in the Top 20 of the most performant AI models according to the Arena, a collaborative benchmark where internet users arbitrate duels between two anonymous models. For the first time in several months, Anthropic ceded the lead in the Text Arena ranking, which evaluates the performance of AI models on tasks such as mathematics, code, and creative writing.

Gemini 4 Argon leads the rankings

Google's Gemini 4 Argon, recently deployed for paying subscribers, took first place, ahead of Meta's Muse Wonder. Anthropic, however, still placed six models in the top part of the table. Gemini 4 Argon, initially reserved for a handful of cyber defense actors, apparently won over the Arena community. Designed for long and complex, multi-step tasks, the model surpasses Muse Wonder, which also claims this specialty.

Anthropic dominates the mid-ranking positions

Behind, Anthropic dominates the top part of the table. Claude Opus 5.5 takes third place, ahead of Claude Fable 5, the "High" version of Claude Opus 4.6 — which remains in the top five positions despite a release dating back to February —, Claude Fable 5.1, Claude Opus 4.7, and the standard version of Claude Opus 4.6. Gemini 3.8 Flash and Muse Spark 1.3 complete the top 10.

OpenAI lags behind major competitors

OpenAI, the company that created ChatGPT, is still far behind: its first model, GPT-5.6 Sol, appears only at 21st place. With the exception of Moonshot, which places Kimi K3 in 18th position, the top 20 is now monopolized by Meta, Google, and Anthropic. Here are the 20 most performant AI models for text-to-text tasks according to Arena: Gemini 4 Argon "High," Muse Wonder "Max," Claude Opus 5.5 "High," Claude Fable 5 "High," Claude Opus 4.6 "High," Claude Fable 5.1 "High," Claude Opus 4.7 "High," Claude Opus 4.6, Gemini 3.8 Flash "High," Muse Spark 1.3 "Max," Claude Opus 4.7, Muse Spark 1.2 "xHigh," Muse Spark 1.1, Claude Opus 5 "Max," Claude Opus 5 "High," Muse Spark, Gemini 3.7 Flash "High," Kimi K3 "Max," Gemini 3.1 Pro Preview, and Gemini 3 Pro.

Arena evaluates models on specialized tasks

Arena also evaluates models on more targeted tasks, such as web search, image analysis, and document analysis. In web search, the "xHigh" version of GPT-5.6 Sol dominates this category, with Qwen 3.8 Max and Claude Opus 4.6 Search completing the podium. In the top ten positions, nine belong to Anthropic (5) and OpenAI (4).

Anthropic leads in image and document analysis

In image analysis, Claude Fable 5 takes the lead, ahead of Qwen 3.8 and Claude Opus 4.6. Anthropic places 5 models in the top 10. In document analysis, Claude Opus 5 comes first in a category where Anthropic monopolizes the top seven places. OpenAI completes the top 10 with GPT-5.5 in standard version, GPT-5.5 "High," and GPT-5.6 Sol.

Arena uses user votes to rank AI models

Arena is scrutinized by the main actors and observers of AI. It is a collaborative benchmark that relies on the votes of internet users to rank the models. The principle involves submitting a request to two anonymized models, which each produce a response and the user selects the most relevant one. Each model is assigned an Elo score, a kind of rating that varies according to its performance in these duels. A victory against a better-ranked opponent earns points, while a loss to a less well-ranked competitor results in losing points. A model at the top of the ranking can always improve its score, but its progress is slower.