What the Vibe Jam 2026 top 50 did differently

Vibe Jam 2026 asked for new, free web games with at least 90% of the code written by AI, and got 945 of them, according to its press page. We list 367 of those entries with the jam’s own final placement, and 34 of them finished in the top 50.

The obvious question is what the top 50 did that the rest didn’t. We lined them up against all 367 entries we list. Four differences stand out, and a fifth check, on the AI tool, shows less than you’d expect.

• A Game About Capybaras Delivering FoodClaude · Simulation
• Fanto's Mega-MartGPT · Action
• WenWareGPT · Puzzle
• Null RangeCursor · Shooter

How we compared

Every number below is live, counted from our catalog each time the site is built, so it moves if we add entries. “Top 50” means the 34 entries we list with a final placement of 50 or better. “Median placement” is the middle rank of a group: half its games placed better, half worse.

Two caveats before the fun part. We list 367 of the 945 entries. And every pattern below is a correlation from one jam, where many things happened at once, so read it as a hint and keep your own judgment.

1. They were 3D, in Three.js

85% of the top 50 run on Three.js, against 68% of all entries we list. The jam’s rules allowed any engine and said Three.js is usually recommended (jam page), and the top of the table took the hint.

The flip side: only 9% of the top 50 state no engine at all, against 19% overall. Engine-free games do reach the top, HALDANE-4 and its ASCII horror among them, but they’re rarer there. Three.js entries have a median placement of #256; everything else, #314.

1. Three.js2985%
2. HTML5 Canvas13%
3. Godot13%

Out of 34 games; a game can count in several bars.

2. They brought friends

Online play opens one of the widest gaps. 65% of the top 50 have online multiplayer, against 46% of all entries. Online games have a median placement of #260; single-player ones, #297.

The jam said it preferred multiplayer games without requiring them (jam page). The podium agrees: all three of the top three we list are online. A game you can send to a friend gets played by two people instead of one, and in a jam that’s how a game gets noticed.

Multiplayer is also the hardest thing to get right with an AI agent. How to make a multiplayer game with AI covers the architecture to ask for.

3. They worked on a phone

76% of the top 50 play on a phone, against 69% overall. The share gap is small, but the median placements sit far apart: #261 for phone-friendly entries, #321 for the rest.

A jam game gets shared as a link, and a lot of links get opened on phones. A game that greets a touchscreen with “press WASD” can lose that player before the first move. The mobile guide covers touch controls and screen fit.

4. Action and simulation beat puzzles

65% of the top 50 are tagged action, against 45% overall. Simulation rises from 17% to 29%: think of the capybara courier balancing a fruit stack, which won the whole jam.

Puzzles go the other way. They make up 11% of all entries and 9% of the top 50, with a median placement of #361. Shooters sit at #222. A puzzle asks the player to sit and think, and a friend who tapped your link in a group chat may not stay that long.

5. The tool mattered less than you’d think

Claude Code is the most-named tool in the top 50, as it is everywhere: 38% of entries in the top 50, against 43% of all entries. Two other numbers are more interesting.

Codex punches above its weight at the very top: it’s named by 32% of the top 50 but only 19% of all entries, and it made two of the top three. Yet its median placement is #290, close to Claude Code’s #279. Cursor has the best median of the three, #209.

1. Claude Code1338%
2. Codex1132%
3. Cursor824%

Out of 34 games; a game can count in several bars.

Read together, the tool shifts where the middle lands a little, but every big tool reached the top. What the top 50 had in common was the game: 3D, online, on a phone, with something physical to do.

What to take into the next jam

If you’re building for the next one, the pattern is short enough to fit on a sticky note:

1. Three.js, in the browser. The jam recommends it, and the top of the table used it.
2. A reason to invite someone. Online entries placed better on median than single-player ones.
3. Touch from day one. Test on your own phone before you post the link.
4. Something physical to do in the first ten seconds. Steer, stack, dodge, throw.
5. Polish the feel. Everyone had AI; not everyone had hit stop and a good jump. See game feel with AI.

Then test it like a player well before the deadline. Playtesting games with AI shows how to let your agent play it first.

Play the top of the table

Every entry we list is on the Vibe Jam 2026 page, where you can play and rate them. Our top five breakdown covers the podium game by game, and the statistics page has the numbers for the whole catalog. The top three won $25,000, $10,000 and $5,000, according to Pieter Levels’s results post.