Anthropic Surpasses Google and OpenAI in 2026 AI Rankings

The 10 Most Powerful AI Models in March 2026

In March 2026, Anthropic continues to dominate the ranking of artificial intelligence models on the LMArena platform. The Claude Opus 4.6 "thinking" model holds the top position, followed by its standard version, Claude Opus 4.6. Google, which had long been at the forefront with Gemini 3 Pro, now sees its models ranked third, fifth, and ninth with the various versions of Gemini.

OpenAI, which had been absent from the top 10 in January and February, makes a notable comeback by placing its models GPT-5.4 and GPT-5.2 in sixth and eighth places, respectively. This performance marks a rebound for OpenAI after several months of withdrawal.

Here are the ten leading language models dominating the market in March 2026, across all tasks:

• Claude Opus 4.6 "thinking"
• Claude Opus 4.6
• Gemini 3.1 Pro
• Grok-4.20
• Gemini 3 Pro
• GPT-5.4 "high"
• Grok-4.20 "reasoning"
• GPT-5.2
• Gemini 3 Flash
• Claude Opus 4.5 "thinking"

The Best AI Models by Category in March 2026

LMArena is not limited to an overall ranking. For several months, it has also offered thematic leaderboards, allowing for a better understanding of each model's capabilities according to the type of task. In text generation, Anthropic maintains its grip with Claude Opus 4.6 and its "thinking" version in the top two spots. Google follows with Gemini 3.1 Pro, while Grok-4.20 and Gemini 3 Pro complete the top 5.

In web development, Anthropic crushes the competition with its iterations of Claude Opus and Claude Sonnet occupying the top five positions. Claude Opus 4.6 is the benchmark in this exercise. Google manages to sneak Gemini 3.1 Pro and Gemini 3 Flash into the top 10, while GPT-5.4 "high" ranks sixth.

For image analysis, Google establishes itself as the reference, with five models in the top 10. Gemini 3 Pro leads the pack, followed by Gemini 3.1 Pro and GPT-5.2. OpenAI successfully places two other models in the upper part of the table.

In image generation, the Nano Banana 2 model, launched on February 26, becomes the default model on the Gemini application and takes the lead in this category, dethroning GPT Image 1.5, which had long been considered the benchmark. Reve 1.5 ranks fifth.

Finally, in web search, the "search" version of Claude Opus takes the top spot in a ranking from which Perplexity models have completely disappeared. Grok-4.20 and the "search" version of GPT-5.2 complete the podium.

What Criteria Underlie the LMArena Ranking?

Founded by students from the University of Berkeley, LMArena (formerly Chatbot Arena) is based on a principle of duels judged by users. In the Battle Mode tab, the platform invites users to submit a prompt, to which two anonymized models respond. The user must then choose between the two results. These confrontations allow each model to be assigned an Elo score, which evolves over time based on the results. A higher-ranked model that loses to a theoretically inferior opponent loses points, while it gains points if it defeats a supposedly stronger model.