LMArena in Detail

[ikaijua]

LMArena in Detail

LMArena is a free AI model voting arena — you prompt, two anonymous models answer, human votes build a live leaderboard. The platform now runs as Arena (arena.ai).

What it is. A crowdsourced LLM evaluation platform born from UC Berkeley's SkyLab, incorporated in 2025 and valued at $1.7B in January 2026.
Signature capability. Anonymous human blind-testing + Bradley-Terry statistics across text, coding, image and voice tracks.
Price. Completely free — the company earns from paid evaluation services for model vendors.
Note. The board reflects human preference, not objective capability; top gaps are usually inside the confidence interval, so watch tiers, not positions.

Floors below follow “Product intro → Background & deep dive → Pricing & caveats”.

Table of Contents

Product intro

1. What is LMArena?
2. How to play: the full voting flow
3. How the leaderboard is computed

Background & deep dive

4. Who's behind it: organizations and founders
5. Timeline
6. How to use it to pick models

Pricing & caveats

7. Pricing & openness
8. Honest caveats: leaderboard limitations
9. Sources

Clicking a TOC item jumps to the floor; if it does not respond in-page, open the link in a new tab.

[ikaijua]

What is LMArena?

One line to remember it. A free "AI model voting arena" — you ask any question, two anonymous AI models answer, you pick which one you like better, and all human votes feed a live leaderboard.

What you can do with it.

• Blind-test models: pit two unknown models against each other on the questions that matter to you
• Check live rankings: separate leaderboards for text, coding, image, voice and more
• Pick a model: get crowd-sourced reputation before buying a subscription or calling an API

What makes it special.

• Real human preference: not benchmarks — the blind choices of millions of users
• Anonymous duels: you vote before the model identities are revealed, so brand names can't bias you
• Free and open: anyone can vote, and model vendors can submit their models to compete

Where to use it. Web at arena.ai (formerly lmarena.ai). No install needed.

How much it costs. Free.

[ikaijua]

How to play: the full voting flow

1. Open arena.ai and enter any prompt (question, code, image description...)
2. Two anonymous models each produce a response, shown side by side
3. Vote: A is better / B is better / tie / both are bad
4. The identities of both models are revealed after you vote
5. Your choice enters the leaderboard; scores update in near real time

Supported modes: text chat, coding, image generation and editing, voice, long context — each has its own independent ranking.

Tip: voting on a variety of difficulties helps calibrate the leaderboard — that is exactly how the platform works.

[ikaijua]

How the leaderboard is computed

Started with chess-style Elo. Originally borrowed from chess: each duel settles by expected win probability — an upset against a strong model gains many points, a strong model beating a weak one gains almost none. A 100-point gap means roughly a 64% win rate for the higher-rated model.

Upgraded to the Bradley-Terry model. Since 2024 the platform uses the Bradley-Terry statistical model (1952): instead of settling duel by duel, it fits a maximum-likelihood model over the entire vote history, deriving the strength values that best explain all observed win rates, then maps them to Elo-style numbers. Two advantages:

• No recency bias — every vote weighs equally regardless of when it happened
• Confidence intervals — the score comes with a stated uncertainty range, so you don't over-read tiny rank differences

Caveat: the leaderboard reflects human preference, not objective capability. Voters favor long answers, polished prose, and brand alignment — see the "Honest caveats" floor.

[ikaijua]

Who's behind it: organizations and founders

• Origin: launched May 2023 by LMSYS, a multi-university research group led from UC Berkeley's SkyLab (with Stanford, UCSD, CMU and MBZUAI), initially as a nonprofit research project
• Core person: Wei-Lin Chiang — CS PhD student at UC Berkeley SkyLab (advised by Ion Stoica), lead of the Chatbot Arena project
• Incorporation: April 2025, spun out as Arena Intelligence Inc., keeping the platform neutral and open; May 2025 raised a $100M seed round (led by Andreessen Horowitz and UC Investments, ~$600M valuation); January 2026 raised a $150M Series A (led by Felicis and UC Investments, with a16z, Kleiner Perkins, Lightspeed and others) at a $1.7B post-money valuation
• Sibling projects: FastChat (training/serving framework), Vicuna, SGLang — also open-source

In one line: from a university lab experiment to a $1.7B independent evaluation company — that is who pays for the cost of running it.

[ikaijua]

Timeline

• May 2023: Chatbot Arena launches as a UC Berkeley academic project
• Mar 2024: LMSYS publishes its mission statement — open, community-driven evaluation
• Sep 2024: renamed LMArena (short for LMSYS Chatbot Arena)
• Apr 2025: incorporated as Arena Intelligence Inc.; new beta site
• May 2025: $100M seed round (a16z + UC Investments)
• Nov 2025: methodology post — Bradley-Terry with confidence intervals and rank spread
• Jan 2026: renamed Arena (domain arena.ai, redirecting lmarena.ai); $150M Series A at $1.7B

One thing to remember: both renames were branding only — the underlying methodology and voting data never changed. Same leaderboard, three names.

[ikaijua]

How to use it to pick models

• Chat/writing models: look at the top of the overall leaderboard, then weigh your actual needs (Chinese, long-form, creativity)
• Coding models: switch to the Coding track and trust the human-voted coding rank over vendor self-tests
• Image models: the image track — Nano Banana, Grok Imagine and others compete there
• Read the confidence interval: if two scores differ by less than the stated uncertainty, treat them as tied — don't pay a premium for a phantom gap
• Re-test yourself: the leaderboard only points a direction; run each finalist on your own workflow for a few days

Pitfall: the top of the board is often statistically indistinguishable — watch the tiers, not the exact positions.

[ikaijua]

Pricing & openness

• Free for users: voting, browsing the leaderboard and blind-testing cost nothing
• Revenue model: the company earns from paid evaluation services for model vendors (evaluation suites, API/enterprise-grade benchmarking); the public leaderboard stays open
• Submitting models: vendors can enter models to compete (some tiers may be paid)
• Data openness: aggregated vote data is partially public for research

For ordinary users: you are always the free player — the platform earns from vendors who want to rank well. That is its key difference from ad-driven sites.

[ikaijua]

Honest caveats: leaderboard limitations

• Preference ≠ capability: humans favor longer, more polished, more confident answers — factual accuracy is not always what wins
• Voter bias: the voter pool skews technical and English-speaking; value for Chinese or niche use cases is limited
• Version confusion: different versions and reasoning tiers of the same model are listed separately — check the right row
• Rank volatility: the board moves weekly; top gaps are usually inside the confidence interval
• Commercialization: paid vendor evaluation services are now a revenue source — the company states it stays neutral and data-transparent, but that is an open public discussion

How to read the board: look at tiers and trends, not positions; your own blind test is the final answer.

[ikaijua]

Sources

• Arena official site & leaderboard: https://arena.ai/leaderboard
• LMSYS About (origin & partner universities): https://www.lmsys.org/about/
• LMArena blog (Hello from LMArena): https://arena.ai/blog/hello-from-lmarena/
• Series A announcement (Felicis + UC Investments, Jan 2026): https://arena.ai/blog/series-a/
• Methodology explainer (Elo vs Bradley-Terry): https://ai-tldr.dev/learn/evaluation-safety/benchmarks-leaderboards/lmarena-elo-ratings-explained/
• Core person Wei-Lin Chiang: https://infwinston.github.io/
• Funding & incorporation summary (Sacra): https://sacra.com/c/lmarena/