Which AI model makes the most fun games?

The short version

• Developers pick Claude Opus. It powered 54% of 4,200 new games on Playgama's AI Game Index.
• Players liked OpenAI Astra games most on that index: a median 65 likes per 1,000 plays, versus 39 for Opus.
• Simon Willison's hands-on take points the other way: Opus 5.5 made games he enjoyed, and he felt GPT-6 Astra's sense of fun wasn't there yet.
• Both can be true. One is a single platform's self-reported data, the other is one experienced developer's games. Neither replaces playtesting your own.

Playgama's AI Game Index: what players liked

Playgama's first AI Game Index tracks 4,200 games published on its Playgama Sandbox between Sept 22 and Oct 21, 2026, and ranks models by share of new games, likes and playtime (GamesBeat, Oct 29). The index updates daily at playgama.com/ai-game-index.

Share of new games

Playgama says Opus climbed to 62% the week before the report, while Sol fell to 8%. Astra peaked at 16% to 17% of daily new games in late September and has stayed under 5% a day since Oct 2.

What players responded to

By agent: the median Codex-built game earned 65 likes per 1,000 plays, versus 43 for Claude Code. The open-source agent OpenCode is behind 1.6% of games and led rated tools on playtime at 5:14.

Read it as a thermometer

Founder Dmitry Kachmar's own words: "Look at the index as a thermometer, not a judge." The caveats that matter:

• It's one platform's games, not the whole market.
• Developers self-report which model and agent they used.
• Figures are medians. A play counts after 30 seconds, a game after 20 plays, a model after 15 measured games.
• Likes show what players reacted to, not overall quality.

Simon Willison's test: Opus 5.5 "knows what fun is"

In The October inflection point for vibe game development (Oct 28), Willison says October 2026 is to vibe-coded games what November 2025 was to coding agents. He built four browser games in HTML and JavaScript without reading the code, and none took more than four hours of his own time.

• He credits Claude Opus 5.5 with real gains in 3D modeling, music and game design.
• His most useful prompt: Make it a LOT more fun.
• He ran the same prompts through GPT-6 Pro: the game "looks great," but he doesn't think it plays nearly as well. His impression is that GPT-6 Astra "isn't quite there yet" on fun, and he expects OpenAI to catch up by the end of the year.
• His workflow: start with "Build a mobile-friendly three.js game where..." in a chat app, then switch to Codex or Claude Code once it's past proof of concept.

So the index's players and Willison disagree on Astra. That's the point: "fun" depends on the game, the players and how much you iterate, so measure it on yours.

Roblox: AI is already in most top games

From Roblox's Q3 2026 earnings (GamesBeat, Oct 29):

• More than 80% of its top 100 games by hours played used AI to accelerate creation in October 2026.
• Time spent in AI tools has tripled among top creators since Q2, and daily use of the Roblox Studio MCP server is up nearly five-fold.
• By the end of Q3, a game created almost entirely with AI broke into the top 10 by concurrent players.
• Its Build tool has been tried by more than 500,000 users, who published over 70,000 games.

How to test fun in your own game

1. Pick a number before you pick a model. Borrow Playgama's: likes per 1,000 plays and median playtime. Even 20 playtesters give you a baseline.
2. Keep the core loop playable in 30 seconds. If a tester can't tell what's fun by then, no model swap will save it.
3. Ask for fun explicitly. Willison's "Make it a LOT more fun" is a real prompt, not a joke. Run it, then playtest again.
4. A/B the model, not just the prompt. Give two models the same spec and let players choose. Your genre may not match the index.
5. Checkpoint every change. "More fun" passes can break the build. Our free Checkpoint Loop runs your checks after every agent step.

Our $19 Vibe Coder Game Jam Starter has an "Is It Fun?" review, a playtest plan and a juice pass built for this loop. Entering the Epic AI MegaJam? See the prep checklist; "fun factor" is one of its seven judging criteria.

See the $19 jam pack

FAQ

Which AI model do developers use most to build games?

On Playgama's AI Game Index (4,200 games published Sept 22 to Oct 21, 2026), Claude Opus powered 54% of new games, followed by OpenAI's Sol at 11% and Claude Sonnet at 8%. Model info is self-reported by developers.

Which AI model makes the games players like most?

On the same index, games built with OpenAI's Astra had the highest median likes: 65 per 1,000 plays, and the longest median playtime at 4:58. Claude Opus games had a median 39 likes per 1,000 plays.

Is Claude Opus 5.5 good at game design?

Simon Willison says yes: in his Oct 28, 2026 tests, Opus 5.5 built four browser games he enjoyed, and telling it to "Make it a LOT more fun" worked. His impression was that GPT-6 Astra could build a game but its sense of fun didn't match Opus 5.5 yet.

How many top Roblox games use AI?

Roblox said in its Q3 2026 earnings that more than 80% of its top 100 games by hours played used AI to accelerate creation in October 2026.