Roblox says 80% of its top games use AI, and a new index tracks which AI builds games

Friday, October 30, brought three stories about AI in games, and they fit together neatly. The biggest one came from Roblox, which used its quarterly results to argue that AI has changed how games get made on its platform, and that it wants AI to help bring games from everywhere else onto Roblox too. A browser games company, Playgama, started publishing a daily index of which AI models people use to build games and how players react to the results. And Studio Atelico opened early access to an engine that runs AI characters directly on the player's device instead of in the cloud.

The common thread: making a game with AI is getting easier. The hard questions are shifting to what happens around the making, from getting noticed to keeping quality up.

Here's what happened, in plain English.

Roblox says AI is now behind most of its biggest games

Roblox is the platform where players, many of them children and teenagers, play games built by other users. On Thursday, October 29, it published its results for the third quarter of 2026, and the shareholder letter that came with them leaned heavily on AI. The company says it has witnessed "an inflection point in AI adoption and content generation."

The numbers it gives to back that up are striking, and they are all the company's own:

• More than 80% of the top 100 games on Roblox, ranked by hours played, used AI tools to speed up creation over the past month.
• Two games, Break and Steal an Egg and Huss Valley, were created almost entirely with AI, and both broke into Roblox's top 20 by concurrent players, meaning the number of people playing at the same moment.
• Daily uploads of new assets to the platform, things like 3D models and items, tripled.

Roblox also has its own AI creation tool, called Build, which it describes as a prompt-to-game tool: you describe the game you want in plain language, and it makes a playable version. Build is still in an early alpha test in just four countries, New Zealand, Singapore, Serbia and Thailand. Even so, Roblox says more than 500,000 people have used it and they have published over 70,000 games with it. GamesBeat reports that the company plans to expand it to more markets by the end of the year.

Creators are also connecting outside AI assistants to Roblox's building software, Roblox Studio. They do that through what's called an MCP server. MCP stands for Model Context Protocol, a common way for AI assistants to plug into other software. GamesBeat explains that Roblox's version lets assistants such as Claude and Cursor work directly with Roblox Studio, and that Roblox is opening it up to every major part of Studio so AI agents can eventually do anything a creator can do by hand. Roblox says the number of people using its MCP server every day grew nearly five times since the second quarter, and that time spent with AI tools among its top creators tripled over the same period.

Roblox's plan: bring any game onto Roblox, and keep the slop down

The most ambitious line in the letter is about something Roblox calls the Universal Player. The company says it includes "the goal of enabling the capability to import any existing console, PC, or mobile game onto Roblox." Put simply, Roblox wants AI to help convert games from other platforms so they can run on its own. As Vertical Slice notes, Roblox pitches this as a way for those games to get its discovery, safety, communication and monetization features. It is a stated goal, not a product anyone can use yet.

The letter also describes Roblox Everywhere, a plan to let creators self-publish standalone games that were made on Roblox. The first phase, publishing a selection of top games, is expected in the first half of next year.

Roblox's founder and CEO, David Baszucki, summed up the company's view in its earnings statement: "AI is fundamentally reshaping the gaming industry." He argued that when anyone can create a game, what matters most is scaled distribution and effective discovery, in other words, getting games in front of the players who will enjoy them. GamesBeat describes Roblox's case this way: building may become a commodity, but an audience, discovery, monetization and safety can't be generated with a few prompts.

More games made more easily also means more low-effort games. Roblox addresses that directly in a section of its letter titled "Surfacing New Content vs. AI Slop." AI slop is the internet's nickname for cheap, low-quality content churned out with AI. According to Vertical Slice, Roblox has adjusted its recommendation and search systems to surface higher-quality experiences and reduce low-quality AI-generated content. The letter says that by the end of the quarter, about 40% of the top 100 games in its weekly recommendations, ranked by impressions, were created in the previous three months, which Roblox presents as a sign that new games can still break through.

The rest of the quarter was mixed. Roblox reported 138 million daily active users, down 9% from a year earlier, while revenue rose 16% to about $1.6 billion. Bookings, the money players spend on the platform in the quarter, fell 12% to $1.7 billion, though Vertical Slice notes Roblox said that still beat the top of its own forecast. Roblox expects daily users to keep falling into the fourth quarter, which it puts down to seasonal patterns. The company has been rolling out age checks for its chat features, which it has said reduced how much players communicate, and it faces lawsuits alleging it failed to protect young users. Investors didn't love the overall picture: GamesBeat reported the stock fell about 11% in after-hours trading, meaning trading after the market's regular close.

A new index ranks AI models by what players actually do

One question keeps coming up as AI-made games multiply: when people build a game with AI, which AI are they using, and does the choice make a difference to the result? Playgama, a company that distributes games played in a web browser, has started trying to answer it with an AI Game Index, reported by PocketGamer.biz.

The index looks at games built with AI tools on Playgama's own platform and, in the company's words, measures them "from observed player behaviour, with no human judges." A game is counted once it reaches 20 plays. Playgama measures how long people play, counting only sessions of 30 seconds or more, and how many likes a game gets per 1,000 plays. Each AI model gets the median score of its games, and a model is only rated once 15 of its games have been measured. As of Friday's update, the index covered more than 5,400 games and more than 1.4 million game launches, using data from September 30 to October 29.

Here is what it shows, and these are Playgama's own numbers. Anthropic's Claude Opus is by far the most common choice, behind about 56% of the AI-built games tracked so far. OpenAI's Sol comes second at about 10%, with Anthropic's Claude Sonnet close behind at about 9%. Google's Gemini models, DeepSeek and xAI's Grok each account for a few percent or less.

The more interesting part is what players do with those games. On likes per 1,000 plays, games made with DeepSeek came out on top at 72, followed by OpenAI's Astra at 65 and Sol at 52. Claude Opus games averaged 38. On play time, the differences are smaller: Gemini Pro games had the longest median session at 4 minutes 30 seconds, with Claude Opus games at about 4 minutes. PocketGamer.biz's summary: the most popular development tool "does not produce the strongest player engagement."

A few caveats are worth keeping in mind. These are browser games on a single platform, so the index says little about the kinds of big-budget games released on consoles. Some models are behind relatively few games, which makes their scores less certain. And the figures move: the PocketGamer.biz article, written a day earlier, cited about 4,200 games and a 54% share for Claude Opus. Playgama says it will update the index daily. Still, it is one of the first public attempts to judge AI-made games by how players actually respond, rather than by how impressive the tools look in a demo.

Studio Atelico wants AI characters to run on your own device

The third story is aimed at developers who want AI-driven characters in their games. Studio Atelico, an independent studio started by veterans of Meta and Creative Assembly, has opened early access to the Atelico AI Engine.

Its main selling point is where the AI runs. Most AI features today work through the cloud: the game sends a request to a powerful computer in a data center and waits for the answer. That costs the studio money every time a player uses the feature, adds a delay, and needs an internet connection. Atelico's engine runs on-device instead, meaning the AI runs on the player's own PC or phone, inside the game.

According to the company, the engine can run language models, image generation, speech recognition and text-to-speech, search, and AI character behavior locally. It plugs into Unity, Unreal Engine and Godot, three of the most widely used game engines, and runs on Windows, macOS, iOS and Linux, on either a graphics chip or a regular processor. The company says that because everything stays on the device, player data doesn't leave it, and studios avoid server costs.

For characters, developers can define personalities, backstories, relationships, motivations and tendencies, and the characters form memories that shape how they act later, so they can change over time instead of following a fixed script. Atelico says it also built in guardrails meant to keep the AI consistent with a game's design. One practical problem with running AI inside a game is that both the AI and the graphics want the same chip. Atelico says a scheduler it built lets the two share it without causing frame drops. That is the company's own claim, and it will be tested once more games use the engine.

The engine already powers Atelico's own debut game, Bobium Brawlers, a turn-based card and dice battler in closed beta. Players describe a creature in plain language and get back a generated portrait and a playable deck. Developers can request access to the engine through Atelico's website.

CEO and co-founder Piero Molino framed the launch as being about tools rather than replacement: "We believe the future of AI in games isn't about replacing developers." He also described the trade-off he says the engine is designed to remove, arguing that until now developers have had to choose between intelligent gameplay, performance and creative control.

The big picture

Put Friday's three stories side by side and a pattern shows up. Roblox says most of its top games already use AI somewhere in their making, and that a couple of hits were made almost entirely with it. Playgama is tracking thousands of games built with AI models in just a month. And tools like Atelico's are trying to move AI from the making of a game into the game itself, in the form of characters that remember and react.

As making games gets easier, the hard part moves elsewhere. Roblox's own pitch is that distribution and discovery are what matter when anyone can build, and it is already tuning its systems to keep low-effort AI games from crowding out the good ones. Playgama's early numbers suggest that the most popular tool isn't necessarily the one whose games players like best. And for AI characters, the open questions are practical ones: whether they run smoothly, whether they stay in character, and whether players actually enjoy them.

None of this settles the wider argument about AI in games, which players and developers are still having loudly. But it does show where the next round of that argument is heading: less about whether AI gets used, and more about what it is used for and whether the results are any good.