ChatGPT vs Gemini vs Claude: which AI builds the best Among Us clone?

Among Us is one of the simplest ideas in gaming: a group of players, a hidden impostor, and a lot of suspicion. That simplicity makes it a perfect test for modern AI coding models. Can they build a full social deduction game from scratch using just a handful of prompts?

Here’s how ChatGPT, Google Gemini, and Claude performed when asked to create an Among Us-style clone from zero—and which one actually produced the best game.

The challenge: build an Among Us clone from scratch

The task given to each AI model was straightforward: create a complete, playable clone of Among Us starting from nothing. The prompt was kept intentionally simple: make an Among Us-style game from scratch, make it a good clone, and first rewrite the prompt to improve it before coding.

Each model followed the same basic flow:

1. Improve the initial prompt into a more detailed spec.
2. Generate the game code (web-based implementation).
3. Fix a small number of issues or bugs (with a strict limit on prompts to keep things fair).

The goal wasn’t to ship a commercial-quality game, but to see which AI could get closest to a fun, working social deduction experience with minimal human guidance.

ChatGPT: solid, mostly bug-free, but a bit bland

ChatGPT (using a high-reasoning 5.5-style model) was up first. After enhancing the prompt, it generated a game called Starfall Protocol, described as an “orbital social deduction” experience.

In terms of features, ChatGPT hit most of the core Among Us beats:

• Lobby with multiple players and one infiltrator
• Simple map with rooms and tasks
• Basic task interactions (e.g., routing power, diagnostics)
• Emergency meetings and a voting system
• AI-controlled crewmates and an impostor

On the plus side, the game was largely bug-free on first run. It launched, players spawned, and the core loop of doing tasks and trying to spot the infiltrator existed. For a single prompt, that’s impressive.

However, there were clear limitations:

• The map was very small and felt cramped.
• The visuals looked generic and “AI-generated,” with placeholder-style art and UI.
• AI behavior and voting logic were sloppy—characters often stood still or behaved strangely.
• The overall experience felt like a prototype rather than a polished clone.

Overall rating: 6/10. ChatGPT delivered a functional base game with the right structure, but it lacked polish, smarter AI, and more interesting level design.

Gemini: fast and flashy, but too many breaking bugs

Next up was Google Gemini 3.5 Flash, configured for high performance. It quickly rewrote the prompt into a detailed spec for a “high fidelity 2D Among Us clone web game” and then started building a game called impostor.io.

Gemini’s strengths showed up immediately:

• Very fast code generation and iteration.
• A more distinct 2D visual style for characters and the map.
• Attempted field-of-view mechanics (limited vision with shadows around corners).
• Customizable player name and impostor count.

But the issues were hard to ignore:

• The game initially wouldn’t even start—the “Start Game” button simply didn’t work.
• After a bug-fix prompt, players spawned but couldn’t leave the first room.
• A second fix finally allowed movement and improved the shadow/vision system, but tasks still couldn’t be used despite pressing the interaction key.
• Kills worked, but the interactions felt glitchy and inconsistent.

In other words, Gemini produced something that looked promising but struggled with basic playability. Even after multiple quick fixes, the game still felt broken at a core interaction level.

Overall rating: 4/10. With more prompts and debugging, Gemini’s version could grow into a strong 2D clone, but under a tight prompt budget it lagged behind due to critical bugs.

Claude: the most complete and feature-rich clone

Finally, Claude Opus 4.8 (on an “Ultra Code” setting) took on the same challenge. While the even more advanced Claude Fable 5 model wasn’t available, Opus 4.8 still had plenty of coding power to work with.

After refining the prompt, Claude generated its own Among Us-style game and shipped a playable build in about 30 minutes. The result immediately felt closer to the original game loop.

Key strengths:

• Clear crewmate role assignment and instructions: you’re a specific color, you have tasks, and there’s an impostor among you.
• A working kill system where the impostor can eliminate players.
• A ghost system after death, letting you continue to move around and observe the game—very true to Among Us.
• AI players that actually moved around the map and participated in meetings and voting.
• Voting logic that produced real outcomes, including mis-votes and ejections.

Visually, the game was still simple and lacked walk animations, but the gameplay loop felt the most complete. You could be killed, watch the impostor hunt others, and see the social deduction logic play out as the AI crew tried (and often failed) to identify the killer.

Overall rating: best of the three. Claude didn’t just match the basic structure—it captured more of the actual feel of Among Us, especially with ghosts and more believable AI behavior.

How the three AIs compare

Looking across all three models, a few patterns stand out:

1. Reliability vs. ambition
• ChatGPT produced the most stable, bug-free experience on the first try, but with simple visuals and basic AI.
• Gemini aimed for more ambitious visual and mechanic choices (like field of view) but stumbled on core functionality.
• Claude struck the best balance: fewer visual flourishes, but a much more complete social deduction loop.

2. Speed vs. quality
• Gemini was the fastest at generating and iterating on code, but speed didn’t translate into a fully working game.
• ChatGPT and Claude took longer, but their outputs were more coherent and playable.

3. Social deduction depth
• Claude’s version felt the most like an actual match of Among Us, with ghosts, visible kills, and messy AI voting that sometimes ejected the wrong player.
• ChatGPT had the right components but weaker AI and smaller scope.
• Gemini’s bugs blocked it from really showing off its social deduction logic.

If you’re interested in broader model strengths and weaknesses beyond this specific challenge, it’s worth checking out a deeper comparison like Claude vs ChatGPT vs Gemini: when each AI model fails (and where it shines), which looks at where each system tends to excel or break down.

What this says about AI coding tools today

This small experiment highlights how far AI coding assistants have come—and where they still struggle.

On the positive side:

• All three models could generate a non-trivial multiplayer-style game from a short prompt.
• They handled game loops, roles, simple AI, and UI without manual boilerplate coding.
• Iterative prompting allowed quick bug fixing and feature tweaks.

But there are clear limitations:

• Even simple games can hide subtle bugs that AIs don’t always catch without explicit feedback.
• Visual polish and UX still require human taste and iteration.
• Complex AI behavior (like convincing social deduction) is hard to get right in one or two prompts.

If you’re considering using these tools for your own projects, it’s less about “which AI is objectively best” and more about which fits your workflow, budget, and tolerance for debugging. For a broader look at that side of things, you might find ChatGPT Plus vs Claude Pro vs Gemini Pro: which $20 AI plan is actually worth it? helpful.

Final verdict: who wins the Among Us challenge?

With a limited number of prompts and minimal human intervention, here’s how the models stack up for this specific task:

• 1st place: Claude – Most complete and feature-rich clone, best social deduction feel, and a working ghost system.
• 2nd place: ChatGPT – Stable, mostly bug-free foundation with all the basics in place, but lacking depth and polish.
• 3rd place: Gemini – Fast and ambitious, but too many critical bugs kept the game from being truly playable.

All three AIs showed they can bootstrap a multiplayer-style game from scratch, but Claude currently offers the most convincing “drop-in” experience for this kind of social deduction project, with ChatGPT close behind as a reliable workhorse and Gemini shining more as a rapid prototyping tool that still needs careful debugging.