Arena

Arena is a web platform where users submit prompts, compare two anonymous model responses in battle mode, and vote for the preferred answer; those votes power public AI leaderboards.

Agent Mode is a product workflow on Arena that lets models autonomously plan and execute multi-step tasks using built-in tools (e.g., web search, coding assistance, file attachments, and a sandbox environment), with a dedicated agent leaderboard built from live behavioral signals.

Agent Arena is the leaderboard methodology Arena uses to rank models based on agentic, real-world traces collected from Agent Mode usage, rather than only single-prompt chat interactions.

AutoEval introduces reward-model-driven leaderboard signals that provide a calibrated “day-1” alternative to waiting for human votes, while still aligning with Arena’s human-evaluation methodology as more votes arrive.

Arena launched a factuality-enhanced ranking for Text and Search Arenas, combining human preference with an additional factuality signal via a weighted scheme and exposing it as an initial non-default toggle in the leaderboard UI.

Arena expanded its coding evaluation beyond frontend prototyping into a fullstack coding platform experience (Fullstack Code Arena), positioned for developers, entrepreneurs, and AI labs to build, deploy, and evaluate production-ready applications.

Multimodal Max is Arena’s model router, extended from text to other modalities, where the router selects among multimodal model variants for direct chat with latency-controlled behavior and expanded capabilities.

Arena added cost and category features to Agent Arena, including cost-per-task views and Pareto-style price-vs-performance exploration, plus category filtering organized around Code, Chat, and Work intents.

Search Arena is an evaluation platform (and dataset/research line) for search-augmented LLM systems, collecting human preference signals on in-the-wild, current-events and real-world search-based interactions.