Chatbot Arena Launches Factuality Leaderboard with Claude Opus 5 Taking Top Spot
Chatbot Arena has launched a new "Factuality" leaderboard that integrates factual correctness into its model rankings, with Claude Opus 5 with Max reasoning securing the top position. The system audits battles by sampling responses, extracting verifiable claims, and checking their correctness head-to-head. This update addresses a major limitation of human-preference benchmarks, which can favor convincing but incorrect answers, by directly measuring factual accuracy. It provides a more reliable metric for developers and users who require high-precision outputs from large language models. The new leaderboard allows users to adjust the weight ratio between preference-based ratings and factual accuracy in real-time. The evaluation process relies on auditing sampled responses to extract and verify claims directly.
## BACKGROUND
LMSYS Chatbot Arena is a popular benchmarking platform that ranks large language models (LLMs) using crowdsourced, anonymous pairwise comparisons and the Elo rating system. While effective at measuring user preference, this methodology historically struggled to account for hallucinations, where models generate incorrect information in a highly confident tone.