Point the Arena at anything that talks - a chatbot, an agent, an API. It plays the customers, the skeptics and the ones who type in all caps. Every answer gets graded by humans and models. You get a score, the failures, and a training set you can fine-tune on.
Personas with goals, moods and bad habits. They push, repeat themselves, change their minds, and try to get your product to say things it shouldn't.
Graders earn a level by agreeing with gold answers; models grade at scale. Disagreements surface the interesting cases first.
Every graded conversation as JSONL - prompt, answer, grade, notes - straight into a fine-tune or an eval set. One request.
An OpenAI or Anthropic model, any HTTP endpoint, or a product with a login - the Arena signs in like a browser.
Personas, topics, rounds. Or let the Arena pick what it hasn't tried yet.
Conversations stream in live. Stop any time. Costs are shown per bout, before and after.
A score, the worst ten conversations, the patterns behind them - and the export.
Free plan: 2 targets, 60 graded conversations a month. No card.
Start a bout