Skip to content
CodeArena

How CodeArena works

Same prompt, one try each, built in advance. You vote blind.

Models
10
Prompts
82
Builds
227
Votes
33

How it works

A blind taste test: two builds from one prompt, you pick, then the names appear.

  1. Step 01

    Read the prompt

    Every pair answers the same prompt, shown word for word. You see exactly what the models saw.

  2. Step 02

    Explore both

    Two builds, labelled A and B, load side by side at the same size. Scroll, click and play with each.

  3. Step 03

    Vote to reveal

    Pick the better build, call a tie or mark both bad. Then see which models made them and what each cost.

Votes feed the leaderboard. After 10 votes, your taste profile unlocks.

How builds are made

Every model gets the same message and one try.

Build a landing page for a rooftop speakeasy cocktail bar - art deco, gold-leaf and emerald, 1920s glamour.Build it now. Output one final ```html block.
A real prompt, as sent. The highlighted line is added to every prompt.
Message
The prompt alone: no system prompt, no tools
Tries
One per model, never redone for quality
Settings
Up to 96K output tokens; each model at its highest effort below max
Built
Sep 24 to Oct 9, 2026, before any vote

82 prompts in six categories, each built by several models, so every pair of models meets on at least one.

The models

Standard list prices per 1M tokens, never discounted rates.

Scoring

Each vote is a match. A Bradley-Terry fit, a close cousin of chess Elo, turns matches into scores.

  • 1000 is average. A 100 point lead wins about 64% of the time.
  • A tie counts as half a win for each side.
  • Models whose score ranges overlap share a rank.
  • A model stays provisional until it has 100 votes and a tight range.

Fairness

Neither build gets a head start.

  • Sides are random, and the A/B split is shown on the leaderboard.
  • Both builds load at the same size, sandboxed, exactly as the model wrote them.
  • Voting unlocks after a few seconds, so you look first.
  • One vote per pair per person. Spam and one-sided voting are not counted.
  • A build that names its own model is removed.

Privacy

No accounts, no names, no emails.

  • An anonymous cookie links your votes. Clear it to start over.
  • IP addresses are hashed, and only used for rate limits.
  • Analytics by Plausible, with no cookies.
  • Builds may load fonts and scripts from public CDNs, like any website.

FAQ

Why are builds made in advance?

Top models take minutes per page (the median here is 4m 44s). Nobody waits that long to vote, so you judge in seconds what took minutes.

Why only one try per model?

That is how most people use them: ask once, then ship or not. Scores pool many prompts, so one off build does not decide a ranking.

How does a new model join?

It builds a set of prompts chosen so it meets every other model, then joins the rotation.

Can I use the data?

The leaderboard is open JSON at /api/leaderboard.

Who runs CodeArena?

Together AI, which serves some of the open models. Every model is prompted, shown and scored the same way. The code is at github.com/Nutlope/codearena.

What happened to V1?

V1 (Nov 2024) generated React components live. Its 6,367 votes were on a different task and different models, so they are not counted here.

Now judge a pair

Two builds, one prompt, names after you vote.

Start judging