The best AI at coding for Roblox Studio right now is a model from Anthropic. It ranks first on Roblox's own OpenGameEval leaderboard (checked October 23, 2026) for both writing new code and fixing broken code, with models from Anthropic and Google close behind on new code. Even the leader passes only about half of the build tasks on its first try, so test everything it writes.
How Roblox measures AI coding
OpenGameEval is Roblox's open-source benchmark for AI assistants in Studio. Roblox announced it on December 17, 2025. It runs each model as an agent in a reproducible Studio environment that simulates editing and playing, then checks the result with executable unit tests. A model only scores if its code works.
The main number is Pass@1: the average chance a model solves a task in a single attempt, averaged over several runs. That is close to how most people use AI, asking once.
The leaderboard is on GitHub, and it lists every model by its exact name. Roblox plans to update it routinely and the page shows no update date, so the scores below are as read on October 23, 2026. AI companies release new models often, and a model that just came out may not have a score yet. For a look at plugins and chat tools instead of models, see our comparison of AI tools for Roblox Studio.
The current leaderboard, by company
Scores below are grouped by the company that makes each model. Open the leaderboard to see the exact model names.
Build set (87 evals), Pass@1
This set tests writing new code.
- Anthropic: 50.34% (1st), 48.05% (tied 2nd), 43.45%
- Google: 48.05% (tied 2nd), 47.82%
- OpenAI (medium reasoning only): 40.69%, 40.23%
Newer isn't always higher here. Anthropic's 43.45% entry is a newer version of the model that scores 48.05%.
Two other columns are worth knowing:
- A Google model has the highest Pass@5, meaning it gets a task right within five tries more often than any other model (63.22%, against 62.07% for the #1 model). If you are happy to retry a few times, it catches up.
- The Anthropic model tied for 2nd has the lowest average tool error rate (0.71%), so it fumbles Studio's tools the least.
Debug set (30 evals), Pass@1
This set tests fixing broken code, which is a lot of what you actually do in Studio.
- Anthropic: 64.67% (1st), 52.67%, 50.67%, 46.00%
- Google: 56.67% (2nd), 51.33%, 49.33%
- Another lab: 56.00% (3rd)
- OpenAI (medium reasoning only): 51.33%, 50.00%, 47.33%
The 2nd and 3rd place models here aren't among the build-set scores listed above. The leader's gap is also bigger on this set, about 8 points over 2nd place.
What Roblox's research paper adds
Roblox published a paper on OpenGameEval on October 1, 2026. It tested 13 frontier models on 84 human-curated tasks, with 16 attempts each. The paper uses a slightly different task set and scoring than the live leaderboard's 87 evals, so the two sets of numbers don't line up exactly.
- The best model solved 51.7% of tasks on a single attempt, but only 39.4% on all five of five attempts.
- No model solved six of the tasks.
- Across all tasks, the top five models were just 1.1 points apart.
- The gaps grow on certain kinds of task: 5.0 points on script-writing tasks and 12.5 points on tasks that change the scene.
In its December 2025 launch post, based on the original 47-task set, Roblox reported that leading models scored near perfectly on simple, single-object tasks, like changing a player's jump power. They struggled when several parts had to work together. Health regeneration and a four-way traffic light got very low scores across every model.
Which one should you pick?
The top models are close, so any of the top few is a fair choice. If you want a simple rule:
- Default: the #1 model on the leaderboard, from Anthropic. It leads both sets.
- You retry a lot: the Google model with the highest Pass@5.
- You want the fewest tool errors in Studio: the Anthropic model with the 0.71% tool error rate.
- You can't use an API key: Roblox Assistant's default model. Roblox doesn't publish which model that is, and it isn't on the leaderboard.
Letting the AI see your place helps a lot
The paper found that runs which inspected every object the correct solution touches passed 13.4 points more often on scene-only tasks and 9.8 points more often on script-only tasks than runs that inspected none. The paper reports this as a pattern it observed, without proving cause and effect. Still, it suggests an AI that can look at your place tends to do better.
There are two ways to give a top model access to your place.
Option 1: Bring your own key to Roblox Assistant
Roblox Assistant is built into Studio. Roblox's docs say it can run and write code, generate materials, meshes and models, simulate virtual inputs, and playtest your game. To choose the model it uses:
1. Open Assistant in Studio. 2. Click the ellipsis (...) and choose Manage API Keys. 3. Add your Anthropic, OpenAI or Google API key.
API keys are paid per use and usually need an adult account (Anthropic requires users to be 18 or older). If you're under 18, ask a parent or guardian, or just use Assistant's default model.
In [April 2026](https://about.roblox.com/newsroom/2026/04/roblox-studio-going-agentic), Roblox announced an improved Planning Mode for Assistant. It reads your code and the data model, asks clarifying questions, and gives you an editable plan before it changes anything.
Option 2: Connect an outside agent with Studio MCP
Studio has a built-in MCP server. Roblox's docs list quick connect for several coding agents and editors, including Cursor, Visual Studio Code and Antigravity, and any client that uses stdio transport can connect through JSON. Through MCP, these tools can read, edit and search your scripts, inspect the data model, run Luau in Edit, Client or Server contexts, start and stop playtests, read console output, take screenshots and send player input. Some clients let you sign in with a subscription you already have, so you may not need a pay-per-use API key.
1. Update Studio. 2. Open Assistant, click the ellipsis (...), choose **Manage MCP Servers**, and turn on **Enable Studio as MCP server**. 3. Connect your client with **Quick connect** in Assistant's MCP server settings, or add a Roblox_Studio entry to your client's mcp.json (examples below). 4. Restart the client. Check for the green indicator under Enable Studio as MCP server, which shows how many clients are connected. If tools don't show up, restart both Studio and the client.
On macOS:
{
"mcpServers": {
"Roblox_Studio": {
"command": "/Applications/RobloxStudio.app/Contents/MacOS/StudioMCP"
}
}
}On Windows:
{
"mcpServers": {
"Roblox_Studio": {
"command": "cmd.exe",
"args": ["/c", "%LOCALAPPDATA%\\Roblox\\mcp.bat"]
}
}
}Roblox warns that MCP clients can read and modify your open places. Only connect clients you trust.
Roblox says 44% of the top 1,000 creators used Roblox Assistant or third-party AI tools through MCP in Studio between March 6 and April 7, 2026.
A workflow that catches AI mistakes
On Roblox's benchmark, the top entries still fail somewhere between a third and half of tasks on the first try, and lower-ranked models fail more often. This routine helps:
1. Ask for a plan first, and fix the plan before any code is written. 2. Let the AI inspect the instances it is about to change. 3. Have it write the code, then start a playtest (or have the agent start one, if it can). 4. Read the console output yourself before you accept the changes. 5. Check that the server makes the important decisions, and that every RemoteEvent handler validates what the client sends.
If your game shows AI-generated text to players, like NPC dialogue from a language model, Roblox [requires](https://create.roblox.com/docs/ai/accelerated-workflows) you to filter it with TextService:FilterStringAsync(), which has to run in a server Script.
Where Picoo fits
Picoo is a paid Studio plugin for building multi-script systems. It reads your open place, so it knows what's already there, and it writes Scripts, LocalScripts and ModuleScripts straight into the right service paths, so you don't copy and paste. It isn't on the OpenGameEval leaderboard, so it has no score there.
- Every script it writes is checked by luau-lsp (type checking with Roblox API definitions) and selene (a Luau linter). Lines that would crash are sent back for a fix before the turn ends.
- Common systems like a combat system, DataStore save, global leaderboard and wave spawner are built from fixed, tested recipes.
- You still press Play yourself. It tells you when to press Play and what to look for.
- Pro is $19/month or $12/month billed yearly (200 credits a month). Business is $49/month or $35/month billed yearly (500 credits a month). Prices are in USD and local tax may apply. There is no free tier and no trial.
If you are learning Luau, Roblox Assistant is free inside Studio and a good place to start, though Roblox doesn't publish its usage limits. Our Assistant vs Picoo comparison covers when each one makes sense.
FAQ
How do OpenAI's models do at Roblox scripting?
They are usable. On the build set they score about 40% (40.69% and 40.23%), against 50.34% for the leader. On the debug set they are mid-pack at 51.33%, 50.00% and 47.33%. The leaderboard only lists them at medium reasoning, and it tests the models as Studio agents, so it doesn't measure the ChatGPT chat app.
Which AI model does Roblox Assistant use?
Roblox doesn't publish which model powers Assistant by default. You can switch it to a model you choose by adding your own Anthropic, OpenAI or Google key under ... > Manage API Keys.
Can an AI write a whole Roblox game for me?
Not reliably yet. In Roblox's December 2025 tests, models handled simple, single-object changes almost perfectly but scored very low on systems where several parts must work together, like health regeneration or a four-way traffic light. In the 2026 paper, six of the 84 tasks were solved by no model.
Is it safe to let an AI agent edit my place?
MCP clients can read and modify any place you have open, so only connect tools you trust. Save a copy of your place first, ask for a plan before changes, and read the console output after each playtest.