人類與 LLM：VibeCodeArena 排行榜帶給我的 AI 編碼啟示

I worked on VibeCodeArena as a consultant with HackerEarth, and while I was busy fixing AI-built apps, I got curious about the leaderboard. This post is about what the numbers taught me.

TL;DR

• A high eval score does not guarantee a top rank. Elo comes from head to head results.
• The top users are scoring higher than the top models, because they work with the AI.
• Always check the confidence and trend, not just the rank.
• Fast is not the same as correct.

The VibeCodeArena models leaderboard (snapshot, 16 Oct 2026)

Quick context: what is VibeCodeArena?

VibeCodeArena is a platform built by HackerEarth where LLMs take on the same coding challenges and go head to head. Every challenge starts with an app that AI has already built. You click Challenge, a sandbox opens, you pick an LLM, and you try to make the app better by prompting or by editing the code yourself. Then you submit, and your version gets evaluated and voted on.

Because everything is evaluated the same way, you end up with one leaderboard for both models and users.

When I started, I was mostly focused on the apps in front of me. Open the challenge, check the preview, read the metrics, write the next prompt. But after a few days I started looking at the leaderboard, and it taught me more about AI than a lot of the articles I had read before.

How the leaderboard works

There are three tabs: All, Models and Users. In the snapshot I looked at, there were 46 models and 100 users. Each entry has:

• Eval score: comes from the evaluation, the same kind of scoring you see in the Metrics tab (security, code quality, correctness, performance and accessibility)
• Elo rating: works like chess. You go up when you win against others and down when you lose
• Win % vs avg: how often that entry beats the average
• Confidence: Low, Medium or High, depending on how much data is behind the rating
• Trend: whether the entry is going up or down

Observation 1: a high eval score does not mean rank 1

This surprised me the most. Here are a few entries from the Models tab:

Claude 3.5 Sonnet has an eval score of 87.5 and sits at rank 3, above Claude Opus 5, which scored 94.0. And one Llama-4-Maverick entry scored 91.3 but is down at rank 9.

So the ranking is not just "who writes the cleanest code". Elo depends on head to head results, which is closer to what people actually prefer when they compare two apps side by side. A model can score well on metrics and still lose matchups because the app does not feel right or misses what the prompt really asked for.

I saw the same thing inside challenges. In the Metrics tab, one app had an overall score of 94% but still had 36 major correctness issues. Numbers are useful, but you have to look underneath them.

Observation 2: the top users are beating the top models

This was my favourite part. On the Users tab, the top users had eval scores around 99.5 and Elo ratings around 1641 to 1642.

Top users have eval scores around 99.5

Put that next to the best model, DeepSeek-v4-Flash, at 95.3 eval and 1639 Elo. In the All tab, where models and users are ranked together, users sit right next to and even above the best model.

In the combined view, users and models are ranked side by side

Why? Because a user's submission is not a human competing against AI. It's a human working with AI. The user starts from what the model built, reads the metrics, notices what is broken, and pushes the model in the right direction. That combination beats the model working alone.

For me this was proof of something I felt during my own work. The LLM writes code really fast, but someone still has to spot the problem, explain it clearly and check that the fix actually works.

Observation 3: confidence matters

A lot of entries had Low confidence, including the rank 1 model. That means there is not much data behind that rating yet, so it can still move a lot. Some lower ranked models had High confidence because they had more matchups.

So don't just read the rank. Check the confidence and the trend. A model with High confidence and an upward trend tells you more than one sitting at the top with Low confidence.

Observation 4: speed is not everything

The model I used a lot in the sandbox was DeepSeek V4 Flash, and it was quick. But fast replies only help when the fix is right. A model that answers in seconds but fixes the wrong thing actually wastes more of your time than a slower one that gets it right the first time.

That's also why the leaderboard mixes so many things together. Being fast, being correct, writing clean code and actually matching the prompt are all different skills.

How this changed the way I prompt

1. I read the Metrics before writing a prompt. The major issues told me where to start.
2. I fixed correctness before looks. A pretty app that breaks the rules of the prompt loses.
3. I re-read the original prompt. Small rules, like "only spend if there is enough balance" in a wallet tracker, are exactly what a model misses and what an evaluator catches.
4. I stopped trusting "fixed". I always checked the Preview again.
5. I treated the model like a fast teammate, not an oracle. Give it clear direction, then review its work.

Final thoughts

Before VibeCodeArena, I mostly saw LLM leaderboards as marketing. Actually working on the platform and looking at the numbers changed that for me. The most interesting thing on the leaderboard is not which model is number one. It's that the best results come from people and models working together.

If you are a student or a fresher, that's a really encouraging thing to see. AI is not making your skills useless. It's making the right skills, like debugging, reading requirements and explaining problems clearly, more valuable.

Have you looked at any LLM leaderboard and noticed something that surprised you? Drop it in the comments, I would love to hear about it.