Arena (formerly Chatbot Arena) vs LM Evaluation Harness
Arena (formerly Chatbot Arena)
arena.aiLM Evaluation Harness
github.comSide by side
| Figure | Arena (formerly Chatbot Arena) | LM Evaluation Harness |
|---|---|---|
| Price from | Free | Free |
| Free plan | Yes | Yes |
| Free trial | — | — |
| Open source | Yes | — |
| Web | Yes | — |
| Linux | — | Yes |
| API | Yes | Yes |
| Self-hosted | Yes | Yes |
| Plans published | 3 | 1 |
| Spec score · LLM Evaluation Tools | 8.8 (#3) | 7.5 (#17) |
Plans
Arena (formerly Chatbot Arena)
- Public ArenaFree
- AI evaluationsPriced on request
- AI EvaluationsPriced on request
In their own words
Arena (formerly Chatbot Arena)
Arena, formerly Chatbot Arena, is a free web platform for evaluating AI models by comparing responses and collecting votes for public leaderboards. In Battle mode, it presents two anonymous responses and reveals the model identities after a vote. Arena aggregates qualifying votes with the Bradley-Terry rating system. Side by Side and Direct modes are also available, but only anonymous Battle votes count toward official rankings. Leaderboards cover agent, text, web development, image, video, vision, document and search tasks; comparisons include image generation, image editing and video generation. Leaderboards can be viewed without signing in, while Direct, Side by Side, Image Arena and Video Arena require login. Accounts can use Google or email, and coding Agent Mode can connect to GitHub repositories. Arena applies a daily usage limit that resets 24 hours after it is reached. Video Arena requires login, permits two generations per 24 hours and offers Battle mode only. Arena also provides commercial evaluation services for enterprises, model labs and developers, with results supported by service-level agreements. Conversations may be used in research datasets and private provider evaluations; shared data is de-identified, and Arena advises against submitting sensitive information.
LM Evaluation Harness
LM Evaluation Harness is a free, open-source framework for evaluating generative language models across many tasks. It lists more than 60 standard academic benchmarks, with hundreds of subtasks and variants. Supported model options include Hugging Face Transformers, GPT-NeoX, Megatron-DeepSpeed, and vLLM. The harness also connects to hosted OpenAI and Anthropic APIs and to local servers compatible with OpenAI APIs. Users can define prompts and metrics, configure tasks with YAML, and extend model backends, filters, metrics, and aggregations through external packages or local modules. Publicly available prompts are intended to support reproducibility and comparison between papers. The base package contains the core framework; model backends are separate optional installs. It runs self-hosted on Linux and API environments, with documented Windows ML and ONNX Runtime accelerator options. MIT licensing applies. APIs without logits or log probabilities are limited to generate-until tasks.