The Pulse
LMArena Raises $200M at a $3.1B Valuation
LMArena has raised $200 million in a Series B round at a $3.1 billion valuation, led by Lightspeed Venture Partners and Khosla Ventures. The company is also launching an Alignment Index as it expands its model evaluations beyond capability
AI.info Team · Written with AI ·
Arena wants its model rankings to serve as an independent measure of AI performance, but the benchmark’s neutrality has faced published criticism. On October 8, the company announced a $200 million Series B at a $3.1 billion valuation, alongside a new index intended to measure whether AI agents act in line with human intent.
The financing gives Arena more resources to expand its paid evaluation business and its public leaderboards. It also sharpens the question raised by critics: how much confidence should users place in a ranking when model developers have a stake in how their systems are tested?
Lightspeed and Khosla lead the $200M round
The Series B was co-led by Lightspeed Venture Partners and Khosla Ventures. Salesforce Ventures, 01 Advisors, Dell Technologies Capital and Endeavor Catalyst also joined, alongside existing investors including Andreessen Horowitz and Felicis, Arena said.
The $3.1 billion valuation marks a steep increase from the $1.7 billion post-money valuation Arena announced for its Series A on January 6. The company says it has now exceeded $100 million in annualized revenue, up from $30 million when it disclosed the Series A. It reached that run rate eight months after launching its enterprise evaluation service, according to a June update.
Arena’s public platform lets people compare model responses and vote on which they prefer. Its commercial service sells model evaluations and performance analytics to AI labs and enterprises, using feedback gathered through the platform. The company says its public-facing site remains free for consumers.
Researchers challenge the leaderboard’s neutrality
The company’s growing evaluation business sits alongside an earlier dispute over how its leaderboard operates. In a 2025 paper, researchers argued that private, pre-release testing let some providers submit multiple model variants and withdraw weaker results, creating a selection advantage. They also raised concerns about differences in how often model providers’ systems appeared in user comparisons.
“Together, these dynamics result in overfitting to Arena-specific dynamics rather than general model quality,” Shivalika Singh and the other authors of “The Leaderboard Illusion” wrote.
Arena disputed parts of the paper’s analysis. In its response, the company said its published policy allowed providers to test multiple variants, subject to capacity, and announced changes including provisional scores in certain cases until more fresh votes had been collected. The disagreement matters to Arena’s pitch: the platform’s value to customers depends on its results being useful beyond the models and providers it evaluates.
Arena adds alignment to capability scores
The Series B announcement also introduced the Arena Alignment Index, which starts with three signals drawn from real agent interactions: unauthorized action, false attribution and deceptive completion. They track whether an agent does something the user did not ask for, assigns a statement or fact to the user against the evidence, or claims to have completed work it did not finish.
Arena says the first release includes methodology and initial results for more than 20 frontier models. The company describes the index as an early measure and says it will add signals over time. “The world needs a neutral third party to measure how safe and aligned AI actually is once it’s in the hands of real people,” Arena said in its funding announcement.
Growth brings a larger test of trust
Arena also reported 350 million sessions across its platform and 62 million votes across text, vision, coding, search, video and image categories since its Series A. It said Agent Arena, its newer space for testing multi-step tasks, had logged 7 million sessions in less than five months.
Those figures show the scale the company can bring to evaluations, while the new index extends its work from which model performs best to how an agent behaves during a task. Arena plans to update the alignment leaderboard signal by signal; the first release covers unauthorized action, false attribution and deceptive completion.