LMArena funding: Every round, investor, and valuation (2026)

LMArena runs the best-known public leaderboard for AI models. Visitors type a prompt, two anonymous models answer, and the visitor votes for the better reply. Those votes rank chatbots, coding assistants, and image and video models, and the AI labs behind them pay the company to test their models before launch.

LMArena has raised $250 million across two rounds since it became a company in April 2025, all of it equity. Its latest round, a $150 million Series A led by Felicis and UC Investments, closed in January 2026. Three weeks later, the company renamed itself Arena and moved to arena.ai, though many people still know it by its old name.

Before it raised venture money, the project lived inside UC Berkeley and ran on grants and donated computing credits. Let's dive into LMArena's funding history.

LMArena's whole venture history fits into eight months. The seed and the Series A came on either side of the launch of its first paid product.

Chatbot Arena went live in April 2023 as a side project of LMSYS, a research group of students and faculty at UC Berkeley. Anastasios Angelopoulos and Wei-Lin Chiang, then PhD students, built it with Ion Stoica, the professor who co-founded Databricks, another Berkeley spinout with a long venture funding history. The site let anyone compare two hidden chatbots, and within months AI labs were citing its rankings in launch posts.

The project paid its bills with outside support rather than revenue. In December 2023, Andreessen Horowitz named the LMSYS team behind Vicuna, FastChat, and Chatbot Arena among the recipients of its open source AI grants, which the firm describes as grant funding rather than an investment. The arena also listed Kaggle, MBZUAI, Together AI, Anyscale, and Hugging Face as sponsors, and none of these amounts were disclosed.

In September 2024, the leaderboard moved to its own domain, lmarena.ai, and separated from LMSYS. In April 2025, it incorporated as Arena Intelligence Inc. A month later, on May 21, Andreessen Horowitz and UC Investments co-led a $100 million seed round, which Bloomberg reported valued the new company at $600 million.

The seed was large for a company with no product to sell. Anjney Midha, the a16z general partner who led the deal, framed the bet as one on reliable, transparent evaluation. The money paid for a rebuilt website, hiring, and research into new evaluation methods.

In September 2025, LMArena launched AI Evaluations, a paid service that runs model tests for labs and enterprises. By December, the company said it was running at $30 million a year in usage-based sales. Felicis and UC Investments then co-led the $150 million Series A, announced on January 6, 2026, and every seed investor named in the round announcement returned.

LMArena's valuation nearly tripled in under eight months. The jump came after the company proved that AI labs would pay for what had been a free academic tool.

The Series A price was about 2.8 times the seed price. The seed figure came from Bloomberg's sources and the company has not confirmed it, while the Series A figure came from LMArena itself. At $1.7 billion, investors paid about 57 times the $30 million run rate the company reported in December 2025.

That multiple has since come down without a new round. In June 2026, Arena said its annualized revenue had passed $100 million, which puts the Series A price at about 17 times current run-rate revenue.

Forge Global, a marketplace for private company shares, lists a September 2026 Series B valuation of $2.88 billion with no amount or investors. Arena has not confirmed it, and its own site still puts total funding at $250 million.

How ownership is split is not public. The founders and the lead investors in each round likely hold the largest stakes. The University of California sits on the cap table both as an investor, through UC Investments, and as the home of the original project.

LMArena's investors fall into two camps: large Silicon Valley venture firms that led or joined both rounds, and Berkeley-linked backers that knew the team from campus. No AI lab is among the disclosed investors, which matters for a company whose product is a neutral ranking of those labs' models.

Andreessen Horowitz co-led the seed and returned for the Series A. The firm knew the team before the company existed through its open source AI grant, which put a16z among the first outside backers of the project.

General partner Anjney Midha said at the seed that AI needs rigorous, transparent evaluation to become reliable. The firm gets an early position in a layer that both model labs and enterprise AI buyers depend on.

UC Investments manages the University of California's endowment and pension assets. It is the only investor to co-lead both rounds.

Chief investment officer Jagdeep Singh Bachher described the seed as a way to turn open AI research into real-world impact. For UC, the deal keeps a share of the value created by a project that started in its own labs.

Felicis joined the seed and co-led the Series A. Peter Deng, a Felicis partner, argued at the Series A that AI progress cannot be measured by lab benchmarks alone.

Lightspeed Venture Partners and Kleiner Perkins invested in both rounds. Lightspeed partner Bucky Moore described Arena as a real-time, community-driven record of how models perform in the wild.

Laude Ventures and The House Fund invested in both rounds, and LDVP joined the Series A. Laude Ventures was co-founded in 2024 by Andy Konwinski, another member of the group behind Databricks' founder-heavy ownership, with $150 million to back University of California spinouts. The House Fund is a venture firm that backs startups from the Berkeley community.

LMArena spends its funding on running the free arena, building paid products, and hiring. The free site is both its data engine and its sales funnel, and it is expensive to operate.

Every vote starts with two model answers, and Arena pays the model providers for most of that inference. Its costs therefore rise with usage. By June 2026, the site had more than 10 million monthly visitors and had logged more than 82 million votes and 700 million conversations.

The company has expanded what the arena can test. After text, it added image input in 2024, then web development, search, image generation, and video. In 2026 it added agent tasks and, in July, a full-stack coding arena.

The money that comes in flows from AI Evaluations. Labs and enterprises pay to test models, often unreleased ones under code names, against real user prompts before launch. OpenAI, Google, and xAI are named customers, and OpenAI tested GPT-5 on the arena under the code name "summit." For a lab, a top ranking is a marketing asset at launch, which is part of how OpenAI sells ChatGPT against rivals with similar models.

Pricing is based on consumption, tied to the volume of battles, votes, and prompts in each campaign rather than seats. That means revenue can swing with the release calendars of a small number of large labs.

Each round announcement named hiring and research as priorities. Contrary Research counted 28 employees in September 2025, a small team for a $1.7 billion company.

Research spending goes into methods that make the rankings harder to game. Arena also publishes parts of its vote data and runs an academic partnership program, which opened a new cycle in September 2026.

Arena is moving from ranking chatbots to ranking agents. Its Agent Mode, launched in mid-2026, logged more than 5 million turns in its first month and was growing about 10% a week, and new leaderboards for agent tasks, automated scoring, and factuality followed over the summer. The commercial goal is to sell to enterprises choosing which model to deploy, not only to the labs that build them.

The main risk is trust. In April 2025, researchers from Cohere Labs and several universities published "The Leaderboard Illusion," which argued that large labs could test many private variants and publish only the best, citing 27 private Meta variants before the Llama 4 launch. Arena rejected the claims, but the same labs whose models it ranks are now its largest customers, which keeps the conflict-of-interest question open.

Competition comes from human-data firms such as Scale AI, whose Meta-backed ownership structure ties it to a single lab, and from Mercor, whose own rapid funding climb came from selling expert data to the same labs. Arena has no IPO plans and has not announced a new round, but with revenue more than three times what it was at the Series A, a larger raise is likely as AI keeps taking a bigger share of startup funding.

LMArena has raised $250 million across two equity rounds: a $100 million seed in May 2025 and a $150 million Series A in January 2026. Before it became a company, the Berkeley project also received undisclosed grants and sponsorships from Andreessen Horowitz, Kaggle, and others. It has not reported any debt.

LMArena's lead investors are Andreessen Horowitz, UC Investments, and Felicis. Lightspeed Venture Partners, Kleiner Perkins, Laude Ventures, The House Fund, and LDVP also invested.

LMArena was valued at $1.7 billion after its Series A in January 2026. Its May 2025 seed round valued it at about $600 million, according to Bloomberg. The company, now called Arena, has not announced a new valuation since.

No. LMArena, now called Arena, is private and has not filed for an IPO. Its shares are not traded on any stock exchange.

LMArena is owned by its founders, employees, and venture investors. Co-founders Anastasios Angelopoulos (chief executive) and Wei-Lin Chiang (chief technology officer) still run the company, with Ion Stoica as a co-founder and adviser. The largest outside holders are likely Andreessen Horowitz, UC Investments, and Felicis, though exact stakes are not public.