Skip to content
TheChatArchiveThe Chat Archive
o3-2025-04-16· OpenAIHugging Face LMArena Human Preference 140K

What's the best AI model currently? I'm talking overall, not for any specific task.

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with Claudeclaude-sonnet-4-20250514-thinking-32k’s answer.
  1. User

    What's the best AI model currently? I'm talking overall, not for any specific task.

  2. Assistant

    “Best” depends on what metric you care about—raw benchmark scores, breadth of skills (text, code, vision, audio), latency + cost, openness, or safety. With that caveat, here’s a snapshot (mid-2024) of the generally acknowledged front-runners and why each could claim the “best overall” crown:

    Rank-ishModel (Release)Org.ModalityContext WindowPublic AccessWhy people call it “the best”
    🥇GPT-4o (May 2024)OpenAIText, images, audio-in/out, video-frames128 k tokens (text) + real-time streaming for speechChatGPT (Plus/Teams/Enterprise)• State-of-the-art reasoning & coding • Near-human latency in voice mode • Consistently tops closed evals (MMLU ~88%, GPQA 53.6%, etc.) • Multimodal in/out in a single model
    🥈Gemini 1.5 Ultra (preview)Google DeepMindText, images, audio, video, code1 M tokens (sparse attention)Limited preview (AI Studio, Vertex)• Longest practical context window (book-length) • Excellent math & code • Integrates natively with search/YouTube/Maps data
    🥉Claude 3 Opus (Mar 2024)AnthropicText, limited images200 k tokensClaude.ai / API• Strong safety alignment • Very good at following long instructions • Top-tier benchmark scores (wins some ARC-E & MATH tasks over GPT-4 Turbo)
    Best open-weightLlama 3 70B (Apr 2024)MetaText8 k (32 k w/ fine-tune)Free to download (permissive licence)• Tops most open LLM leaderboards • Fine-tunable for custom use • Runs on a high-end single GPU
    Dark-horsexAI Grok-1.5V (Apr 2024)xAIText + images128 kX (Twitter) Premium• Multimodal, open-weight soon • Strong code & reasoning, quick iterative rollout

    Key takeaways and how to decide “best”

    1. Raw capability
      – On practically every public benchmark that matters (MMLU, GSM-8k, HumanEval, etc.), GPT-4/Turbo/O still nudges out rivals by a few points.

    2. Multimodality
      – GPT-4o and Gemini 1.5 Ultra are the only single-checkpoint models that do text, vision and audio at parity. If “overall” = “handles any medium,” these win.

    3. Context length vs. working memory
      – Gemini’s 1 M tokens lets you paste several research papers or whole codebases at once. If that’s crucial, Gemini is “best.” GPT-4o is faster but limited to 128 k.

    4. Openness & self-hosting
      – Want source weights, local inference, no vendor lock-in? Llama 3 70B or Mistral’s Mixtral 8x22B are the best trade-offs today. They’re ~1 generation behind the closed champs but rapidly narrowing the gap.

    5. Safety / alignment
      – Anthropic’s Claude lineage stakes its reputation on constitutional AI and refusing unsafe requests while still reasoning well.

    6. Price & latency
      – GPT-4o costs 5×–10× less per token than the original GPT-4, and voice latency is ~300 ms—making it the most practical “best” for production right now.

    So … a one-liner answer • If you just want the single strongest all-around model you can hit with an API today, pick GPT-4o.
    • If you need book-length context or deep Google ecosystem integration, pick Gemini 1.5 Ultra.
    • If you need open weights, pick Llama 3 70B (or Mixtral 8x22B if you prefer Mixture-of-Experts).

    Everything else is nuances of cost, policy, and personal taste.

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