For the first time in several months, Anthropic cedes the top spot in the Text Arena ranking, which evaluates AI model performance in real-time on tasks like mathematics, coding, or creative writing. Gemini 4 Argon, deployed by Google to its paying subscribers, takes first place ahead of Muse Wonder, unveiled by Meta, while the creator of Claude still places six models in the top 10.
Introduced in early October and initially reserved for a handful of cybersecurity players, Gemini 4 Argon has clearly won over the Arena community, a collaborative benchmark where internet users arbitrate duels between two anonymous models. Designed for long and complex, multi-step tasks, Mountain View’s most advanced model edges out Muse Wonder, which also claims this specialty.
The 20 most powerful models
Behind the leading duo, Anthropic occupies most of the top of the chart. Claude Opus 5.5 takes third place, ahead of Claude Fable 5, the ‘High’ version of Claude Opus 4.6, Claude Fable 5.1, Claude Opus 4.7, and the standard version of Claude Opus 4.6. The latter remains in the top five positions despite a release dating back to February. Gemini 3.8 Flash and Muse Spark 1.3 complete the top 10.
OpenAI, on the other hand, remains in the background. Its first model, GPT-5.6 Sol, only appears in 21st place. With the exception of Kimi K3, placed 18th by Moonshot, the top 20 is composed of models from Meta, Google, and Anthropic.
The best models by category
Arena also evaluates models on more targeted tasks, such as web search, image analysis, or document analysis. In web search, the ‘xHigh’ version of GPT-5.6 Sol comes out on top, followed by Qwen 3.8 Max and Claude Opus 4.6 Search. Of the top ten spots in this category, nine go to Anthropic, which gets five, and OpenAI, which gets four.
In image analysis, Claude Fable 5 takes the lead, ahead of Qwen 3.8 and Claude Opus 4.6. Anthropic places five models in the top 10. In document analysis, Claude Opus 5 comes out on top, and Anthropic occupies the top seven spots. OpenAI completes the top 10 with GPT-5.5 in standard version, GPT-5.5 ‘High’, and GPT-5.6 Sol.
How the ranking works
Closely watched by major AI players and observers, Arena relies on votes from internet users to rank the models. A query is submitted to two anonymized models, each of which produces a response, and the user designates the most relevant one. Each model receives an Elo score, a kind of rating that varies based on its results in these duels.
A win against a higher-ranked opponent earns points, while a loss to a lower-ranked competitor causes a loss of points. A top-ranked model can continue to improve its score, but its progress is slower.
