Top AI Models of June 2025: Discover the Complete Performance Rankings!

Are we on the brink of a new era in the AI race? Recently, Google has stepped up its game. The company, established in 1998 and after a rocky start, is now coming off a strong period highlighted by enhancements to its API, the release of a new lightweight model, and a partnership with Canva for video. This indicates that Google is maintaining a strong presence across all fronts of AI. This month, the company continues to dominate the Chatbot Arena, which it took over two weeks ago. However, our traditional model rankings are undergoing some changes!

The Chatbot Arena evolves into LMArena

Instead of one, the Chatbot Arena will now feature seven distinct rankings. Previously, a dynamic leaderboard ranked general models. However, this did not consider the diversity of AI capabilities. While the general ranking still exists (see featured image), it is now joined by six more specific ones:

• Text: this ranking evaluates models’ abilities to understand and generate text, measuring their performance in writing, reformulating, and responding.
• Web Development: this ranking assesses models’ skills in writing, debugging, or explaining web code.
• Image Analysis: this ranking measures models’ abilities to understand and describe images.
• Online Search: this ranking compares models on their ability to simulate an information search and provide a reliable summary.
• Copilot: this ranking evaluates the relevance of models in coding assistance, particularly in completing or generating functions.
• Image Generator: this ranking assesses models’ abilities to create images from textual descriptions.

To reflect this new direction, the platform has been renamed to LMArena.

AI: Top Performing Models by Domain

The overall winner of the rankings is Gemini-2.5-pro-preview, which appears to excel in coding, mathematics, and creative writing alike. In second place, OpenAI’s model o3 also displays excellent performance but seems to falter with longer requests. It is not until the sixth position that we see a model not owned by Google or OpenAI, with Claude-Opus-4. Clearly, the new format of the LMArena is benefiting Anthropic, which had been languishing at the lower end of the rankings for several months.

Text: Top 10 AI Models

In the text category, the top three is dominated by Gemini and ChatGPT, who seem to share the leading positions unsurprisingly. The score gaps remain minimal, indicating that the battle is fought over nuanced, sometimes subjective details.

Web Development: Top 10 AI Models

In web development, the rankings have slightly shifted with the addition of DeepSeek and several Claude models. However, Gemini continues to lead the pack.

Image Analysis: Top 10 AI Models

In visual tasks, Google and OpenAI face no competition. The top 10 is completely dominated by their models. This ranking is telling and reveals the blind spots of the previous version of the LMArena.

Online Search: Top 10 AI Models

The ranking dedicated to online search confirms the rise of specialized models like those from Perplexity, which is very present in the top five thanks to its Sonar answer engine. However, Google, with its model Gemini-2.5-grounding, takes the lead. OpenAI, on the other hand, falls behind in this area, with API versions of GPT-4o remaining secondary.

Code Assistants: Top 10 AI Models

In the code assistant functions, a variety of players such as DeepSeek, Claude, Qwen, Mistral, and Gemini vie for the top position. Unlike previous rankings, no giant manages to establish clear dominance here.

Image Generation: Top 10 AI Models

In image generation, OpenAI makes a significant breakthrough with GPT-Image-1, which outperforms Google’s models. Behind them, a slew of specialized competitors like Ideogram, Recraft, or Stable Diffusion are observed.

LMArena: Ranking Criteria

The LMArena is an open-source platform designed to evaluate artificial intelligence models based on anonymous comparisons made by users. Participants are invited to choose, for each prompt, which of the two presented models provided the best response. This duel system helps to limit biases and ensures a more balanced evaluation of performances.

Each confrontation affects an Elo score assigned to each model. This score, borrowed from the chess world, increases when a model wins against a higher-ranked opponent, and decreases when it loses to a perceived weaker competitor. Now, the LMArena is no longer limited to a global ranking: it offers six distinct tables, based on the same system as originally.