Skip to content
The Chat Archive
Perplexitysonar· PerplexityHugging Face LMArena Search Arena 24K

lmarena top 3

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with gpt-4o-search-preview-high-loc’s answer.
  1. User

    lmarena top 3

  2. Perplexity Assistant

    As of April 2025, the top 3 models on LMArena (based on available data) are dominated by cutting-edge releases from leading AI organizations, though specific rankings depend on the benchmark category and recent updates:


    Overall Leaderboard Highlights

    1. Google Gemini 2.5 Pro

      • Key Strength: Leads in human preference benchmarks (e.g., LMArena Elo) and scientific reasoning (84% on GPQA Diamond)[7].
      • Performance: Achieved a significant margin over competitors in coding (SWE-Bench Verified: 63.8%) and math/science tasks (AIME 2025: 86.7%)[7].
    2. Meta Llama-4-Maverick-Experimental

      • Note: A customized, non-public variant optimized for human voting (Elo ~1417) sparked controversy for potential leaderboard manipulation[8].
      • Public Version: The official Llama 4 release (available on Hugging Face) may rank lower due to less verbose, emoji-free responses[8].
    3. DeepSeek-V3-0324

      • Metrics: Holds an Arena Elo of 1370 (per OpenLM’s April 2025 data)[4].
      • License: MIT-licensed, making it a top open-source contender[4].

    Specialized Leaderboards

    • Text-to-Image:
      • Imagen 3 002 (Google) ranked #1 as of January 2025[6].
    • Coding:
      • Claude 3.7 Sonnet (70.3% on SWE-Bench Verified) and Grok 3 Beta (93.3% on AIME 2025 with multiple attempts) show strong performance[7].

    Key Observations

    • Controversy: Meta’s experimental Llama 4 variant raised transparency concerns, prompting LMArena to revise leaderboard policies[8].
    • Competitive Landscape: Google, Meta, and DeepSeek lead, but xAI’s Grok 3 and Anthropic’s Claude 3.7 are rising challengers[5][7].
    • Open Models: DeepSeek-V3 and QwQ-32B (Apache 2.0) represent high-performing open-source options[4].

    For the latest rankings, visit LMArena’s live leaderboard.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.