AI model rankings and comparison

No single model is best at everything. Rankings change with the task, and they change every few weeks. This page gathers the latest figures from independent leaderboards, each with its date and source, and we update it regularly.

Which one for your use?

• General chat and writing: Gemini 4 Carbon and Gemini 4 Argon lead, with Claude Opus 5.5 very close. The top models are only a few points apart.
• Coding and building sites and apps: Claude models (Fable 5.2, then Opus 5.5) lead the WebDev board by a clear margin.
• Web search: OpenAI's GPT models lead the search board.
• Understanding images and scanned documents: Claude models lead the vision board by small margins.
• Cost: List price is not enough: models that think for longer use more tokens per task. Work out cost per task, not price per million tokens.

General chat and writing (Arena Text)

Measured: 2026-11-11

Building websites and apps (Arena WebDev)

Measured: 2026-11-12

Web search (Arena Search)

Measured: 2026-11-09

Understanding images (Arena Vision)

Measured: 2026-11-02

Composite intelligence index (Artificial Analysis v5)

Measured: 2026-10-28

How to read these numbers

• Arena boards come from blind human votes between two answers. They measure preference, not absolute accuracy.
• Voters select themselves, so results lean toward the tastes and tasks of that platform's users.
• A new model needs weeks before its rank settles, and gaps under roughly 10 points mean little.
• Composite indexes such as Artificial Analysis combine several tests that may not resemble your company's work.
• The sound decision: try two or three models on one real task of yours, with the same data, then compare.