In a recent announcement on Thursday, Arena, which began as a UC Berkeley research initiative in 2023 focused on crowdsourced AI model rankings, has successfully closed a Series B funding round, raising $200 million at a valuation of $3.1 billion, according to their statement.
This significant investment follows the company’s prior disclosure in June that it had achieved $100 million in annualized run-rate revenue.
The funding round was spearheaded by Lightspeed Venture Partners and Khosla Ventures, with additional participation from Salesforce Ventures, 01 Advisors, Dell Technologies Capital, Endeavor Catalyst, a16z, Felicis, and others. Arena had previously announced a $150 million Series A round in January, which valued the company at $1.7 billion post-money. At that time, its annualized revenue was reportedly $30 million. This indicates that the company’s valuation has nearly doubled in approximately ten months.
Arena offers a crowdsourcing platform that consumers can utilize without charge. Users submit prompts or request projects with specific ‘vibe-codes’ and then evaluate which AI model performs best. Arena reports attracting tens of millions of visitors each month.
Last September, the company unveiled its commercial offering, AI Evaluations. This service provides AI model developers and enterprises with in-depth performance analytics derived from its community feedback. The timing for this launch proved ideal, as AI labs began to recognize this year that their models were exploiting benchmarking tests, achieving high scores without genuine improvement. Simultaneously, businesses sought assistance in identifying which models best suited their specific internal requirements, rather than solely relying on standard benchmarks.
“AI is progressing at a pace that outstrips our current evaluation capabilities, and fixed benchmarks become ineffective once models learn they are being assessed,” the company stated in its funding announcement. “The world requires an impartial third party to objectively measure the safety and alignment of AI once it is in the hands of real users. Arena is now fulfilling that critical role,” it added.
In line with this mission, Arena has also introduced a new category to its leaderboard: alignment. This category assesses and ranks models based on criteria such as unauthorized actions (performing tasks not requested); false attribution (incorrectly crediting statements or facts); and what it terms “deceptive completion” (falsely claiming to have completed tasks it did not).
Currently, a selection of OpenAI’s models occupy the top positions on its initial alignment leaderboard, with Claude Opus 5.5 and Claude Fable ranking sixth and ninth, respectively.