チャットボット比較

ChatBot Comparison (Arena, formerly LMArena) compared with Artificial Analysis and Hugging Face Open LLM Leaderboard
ChatBot Comparison (Arena)	Artificial Analysis	Hugging Face Open LLM Leaderboard
Evaluation method: Rankings come from community votes on anonymous model pairs. A Bradley-Terry model, similar to Elo, turns the votes into ratings, with a style control factor. Only votes cast while model names are hidden count.	Evaluation method: It runs its own benchmarks and measurements. Its methodology covers intelligence, quality, performance and price for models, API endpoints and systems.	Evaluation method: Models were scored on six automated tests: IFEval, BBH, MATH, GPQA, MUSR and MMLU-PRO. No human preference is involved.
What is ranked: Leaderboards cover text, WebDev, vision, document, search, image and video generation, among others. The ranking reflects human preference, so it can favour answer style.	What is ranked: Includes language models, speech, image, video and music models. It also measures inference speed, time to first token, and price per token and per task.	What is ranked: Open-weight language models only, evaluated under a standardised protocol so results can be compared and reproduced.
Measurement limits: Preference votes are subjective and depend on who votes and which prompts they write. It does not report latency or price in the way an engineering benchmark does.	Measurement limits: Speed results use a common token unit, and reasoning token counts come from a set of 60 prompts. The page does not address independence or funding.	Measurement limits: The Space shown as running in the cached view, while a secondary source states that the leaderboard has since been archived. Check its current status before relying on it.
Cost and openness: The leaderboards can be viewed freely. Voting needs no purchase.	Cost and openness: Reading the published indices is free of charge on the site. Methodology pages describe how each index is built.	Cost and openness: Hosted by Hugging Face as a community Space, with results in public datasets.
いつ選択すべきか： Choose it to see which chatbot people prefer for writing, coding or general chat, and to follow new releases.	いつ選択すべきか： Choose it when speed, price and measured intelligence must be compared across providers for an engineering decision.	いつ選択すべきか： Choose it for historical comparison of open-weight models on fixed benchmarks, mindful of its possible archived status.

ChatBot Comparison (Arena, formerly LMArena) compared with Artificial Analysis and Hugging Face Open LLM Leaderboard

ChatBot Comparison (Arena)

Artificial Analysis

Hugging Face Open LLM Leaderboard

Evaluation method: Rankings come from community votes on anonymous model pairs. A Bradley-Terry model, similar to Elo, turns the votes into ratings, with a style control factor. Only votes cast while model names are hidden count.

Evaluation method: It runs its own benchmarks and measurements. Its methodology covers intelligence, quality, performance and price for models, API endpoints and systems.

Evaluation method: Models were scored on six automated tests: IFEval, BBH, MATH, GPQA, MUSR and MMLU-PRO. No human preference is involved.

What is ranked: Leaderboards cover text, WebDev, vision, document, search, image and video generation, among others. The ranking reflects human preference, so it can favour answer style.

What is ranked: Includes language models, speech, image, video and music models. It also measures inference speed, time to first token, and price per token and per task.

What is ranked: Open-weight language models only, evaluated under a standardised protocol so results can be compared and reproduced.

Measurement limits: Preference votes are subjective and depend on who votes and which prompts they write. It does not report latency or price in the way an engineering benchmark does.

Measurement limits: Speed results use a common token unit, and reasoning token counts come from a set of 60 prompts. The page does not address independence or funding.

Measurement limits: The Space shown as running in the cached view, while a secondary source states that the leaderboard has since been archived. Check its current status before relying on it.

Cost and openness: The leaderboards can be viewed freely. Voting needs no purchase.

Cost and openness: Reading the published indices is free of charge on the site. Methodology pages describe how each index is built.

Cost and openness: Hosted by Hugging Face as a community Space, with results in public datasets.

いつ選択すべきか： Choose it to see which chatbot people prefer for writing, coding or general chat, and to follow new releases.

いつ選択すべきか： Choose it when speed, price and measured intelligence must be compared across providers for an engineering decision.

いつ選択すべきか： Choose it for historical comparison of open-weight models on fixed benchmarks, mindful of its possible archived status.