LMArena

LMArena — launched in April 2023 as Chatbot Arena, renamed LMArena in 2024 and rebranded as Arena in January 2026 — is a public, web-based platform that evaluates large language models through anonymous head-to-head comparisons. A user submits a prompt, receives answers from two unidentified models, and votes for the better response, after which the models are revealed. The resulting leaderboards have become among the most widely cited in the AI industry, used by developers, journalists and model makers themselves.^[1][2]

History

Chatbot Arena was released on 24 April 2023 by members of LMSYS Org, a research collective, together with UC Berkeley's SkyLab. Its open-source, live approach contrasted with static academic benchmarks, and a March 2024 paper describing the platform reported that it had already amassed over 240,000 votes and that crowdsourced votes agreed closely with those of expert raters.^[2]

The platform expanded steadily: image support followed in June 2024, and in September 2024 it moved to its own dedicated domain, lmarena.ai. In April 2025 the project incorporated as an independent company, and in May 2025 it raised a US$100 million seed round at a valuation of US$600 million. A further US$150 million Series A in January 2026 valued the company at about US$1.7 billion. On 28 January 2026 the platform was rebranded simply as Arena, having added video generation to its arenas earlier that month.^[1][3][4][5]

Key Concepts and Technology

The platform's core is the anonymous "battle": two models answer the same prompt, the user votes without knowing which model produced which answer, and votes accumulate into leaderboards computed with statistical rating methods. Users can also compare chosen models directly, and separate arenas cover text, image understanding, image generation, video, web development, search and agentic tasks.^[1]

Because voting is open to anyone, the platform captures preferences from a very large and diverse pool of users rather than a small panel of experts.^[2] Its rankings have also influenced model releases: companies have previewed upcoming models anonymously on the platform, including DeepSeek, whose prototypes appeared months before its reasoning models drew wide attention, OpenAI's GPT-5 under the codename "summit", and Google's image model that became known as Nano Banana.^[1] The methodology has attracted criticism as well — researchers have shown that organised voting can skew rankings, and commentators have questioned how well preference votes measure capability; the platform tightened its policies in 2025 after a controversy over a model version that differed from its public release.^[1][6]

Applications and Impact

For users, LMArena offers a practical, continuously updated way to compare models on real questions rather than fixed test sets, and its leaderboards are frequently cited in press coverage and model announcements. Its rise reflects a broader shift in evaluation, in which live human preference data is treated as a complement — and sometimes a corrective — to academic benchmarks such as MMLU and SWE-bench.^[1][2]

>See Also

• AI Benchmarking
• SWE-bench
• LLM-as-a-Judge
• Large Language Models
• Nano Banana