AI benchmarks are an important part of how the industry measures progress, and they have played a valuable role over the years. However, as AI evolves rapidly, many factors can affect how useful a benchmark is. Different benchmarks measure different things, and the landscape is always changing. It is important to consider a wide range of perspectives, since no single benchmark can capture everything.
DualMind Arena · Early access
Two models.One prompt.You decide.
DualMind Arena runs blind battles between AI models. Read two anonymous answers to the same prompt, vote for the better one, then see who wrote them. Every vote moves a public ELO leaderboard.
We'll only email you about the launch. No spam, and you can opt out anytime.
Reading the live leaderboard…
In the arena now
53 modelsWorkers AIGroqGoogle
- llama-3.3-70b-instruct-fp8-fast
- llama-3.1-70b-instruct
- deepseek-r1-distill-qwen-32b
- qwq-32b
- qwen2.5-coder-32b-instruct
- qwen3-30b-a3b-fp8
- nemotron-3-120b-a12b
- mistral-small-3.1-24b-instruct
01Why it exists
Benchmarks are broken. Human judgment isn’t.
Static benchmarks leak into training data and stop telling models apart. Blind, side-by-side votes from people asking real questions hold up longer.
- 01
Blind battles
Two anonymous models answer the same prompt. You judge the words, not the logo: the names appear only after you vote.
Model A · Model B
- 02
Your vote is the benchmark
Pick A, B, a tie or both bad. Each vote updates ELO ratings that weigh who a model beat, so one lucky win doesn’t crown a champion.
ELO · the chess rating system
- 03
A public leaderboard
The rankings and the API behind them are public. Read the leaderboard, or query it yourself, without an account.
GET /api/arena/model-stats
02Try a battle
Read both. Pick one.
This is a battle in miniature. Read both answers, then vote. In the Arena, the names appear after your vote; here, you find out what each answer was doing.
Battle · demoPrompt 1 / 3
You asked
Why do AI benchmarks stop being useful?
Three reasons. Test questions leak into training data, so a model can score well by memory. Top models crowd the ceiling, so the gaps between them stop meaning much. And once a score becomes the target, teams optimise for the test instead of the skill. Fresh tasks and blind human comparisons hold up longer.
03Live from the Arena
The leaderboard, right now.
Read from the Arena’s public API as this page loaded. Every model starts at 1000 ELO; beating a stronger model earns more points than beating a weaker one.
Top 5 by ELOELO
| Rank | Model, provider, battles and win rate | ELO |
|---|
Source: api.dualmindlab.tech/api/arena/model-statsHow ELO works(opens in a new tab)API reference(opens in a new tab)
Be there at launch.
One field. Takes five seconds.