# The best AI for Godot: we ran 28 models on 15 Godot 4 tasks

"AI for Godot" is two questions. Which model writes Godot 4 code that actually runs, and which kind of tool puts that model to work: a coding assistant, an editor plugin, an agent over MCP, or an app that builds the game. We answered the first one by running it: 28 models, 15 tasks, every file executed headless on Godot 4.7.2.

Source: https://flockbay.com/blog/best-ai-for-godot/ · Updated 2026-10-18

## Most models now pass, and the cheap ones are not the ones that fail

In 2024 the standard complaint about AI and Godot was that models wrote Godot 3: yield instead of await, KinematicBody2D instead of CharacterBody2D, export var instead of @export. We wanted to know whether that is still true, so we stopped asking and measured.

Each of the 15 tasks asks for one GDScript file with a stated API: move a CharacterBody2D, save and load a dictionary as JSON, paint cells on a TileMapLayer, cast a physics ray, find a path on a grid, wait for a signal or a timeout. A check scene loads the file and drives it on Godot 4.7.2, headless, at a fixed 60 frames per second, and compares the results. A model that fails gets Godot's own output back and one chance to fix it.

Twelve of 28 models passed all 15 on the first try, including three of the five cheapest hosted models we measured: DeepSeek V4.1 Flash, GPT-6 Luna and GPT-5.6 Luna. The expensive models did not separate themselves from them on these tasks. The spread is at the bottom, where five models failed four or more tasks on the first try and one passed only two.

Four of the 28 ran as coding agents (Claude Code and Codex) that could read and write files; the other 24 were asked through their API with no tools, one request per attempt, the way Flockbay reaches its hosted models. The agents had no clear advantage on these tasks: Claude Opus 5.5, Claude Sonnet 5.5 and GPT-6.1 Sol scored 15 of 15, like nine API-only models, and Claude Haiku 4.5 as an agent scored 11.

## The tasks that break models are timing, not syntax

Godot 3 syntax has mostly gone. Across 27 models (leaving out the outlier), the task that failed most on the first try was waiting for a signal or a timeout, whichever comes first: 13 models got it wrong, most often returning false for a signal that arrived well inside the timeout, or never returning at all. Building and playing an animation in code came next, with 10 failures, nearly all the same mistake: the track was built so the sprite never blinked.

A path on a grid around walls failed 6 times, once because GDScript refused a standalone lambda. Health with signals failed 3 times by emitting died twice. The Godot 3 habits are now the rare failures, and they are exactly the ones we wrote up: FileAccess.MODE_WRITE that does not exist, [JSON.parse() called as a static function](/godot/how-to-parse-json-in-godot-4/), and [a := declaration inferred from a Variant](/godot/the-variable-type-is-being-inferred-from-a-variant-value/).

That shifts what to check when an AI writes your Godot code. Read the timing: awaits, timers, signals that can fire twice, animation tracks. That is where a plausible script does the wrong thing without an error.

## Then pick the kind of tool by how much code you want to own

A model is only half the answer. The other half is what sits between it and your project. For the full roundup, with a price and a verdict on each tool, read [the best AI tools for Godot](/blog/best-ai-tools-for-godot/); for choosing a model in depth, [the best LLM for Godot](/blog/best-llm-for-godot/).

If you write code and want help: a general coding assistant such as Copilot, Cursor or Claude Code, pointed at your project folder. You stay the programmer; it drafts and fixes. Tell it your Godot version and paste errors in full.

If you want the AI inside the editor: plugins see the scene and script you have open. AI Assistant Hub is free and open source and can use a local model; Ziva is a paid plugin with an agent that edits scenes and scripts.

If you want an agent that can drive the editor: an MCP server for Godot lets an agent such as Claude Code or Codex launch the editor, edit scenes, run the game and read its errors. That is powerful for developers, and it still leaves you to judge what the agent built.

If you want the game built for you: Flockbay is a free desktop app for Windows, macOS and Linux with Godot 4.7.2 inside. Its AI writes the GDScript, runs the game headless to check its work, the same way the tests above did, and plays it; you play it and judge it, and the project is an ordinary Godot project you can open in the editor whenever you like. It runs on its hosted models or on your own Claude, ChatGPT or agent.

## What to try this evening

Pick one task from the list above that your game actually needs, a save system or a timer that can be cancelled, and give it to the AI you are considering. Then run it, and test the timing on purpose: fire the signal early, fire it twice, quit halfway.

One evening on a small project will tell you more than any comparison, including this one. The tests above are public on each model's page, with the code it wrote and Godot's output, so you can see exactly where a model went wrong.

## Every model on the same 15 Godot 4 tasks, run headless on Godot 4.7.2, sorted by first-try passes.

Read from each product's own pages on 2026-10-18.

| Model | First try | After one fix | How it ran |
| --- | --- | --- | --- |
| Claude Opus 5.5 | 15 of 15 | 15 of 15 | Claude Code agent |
| Claude Sonnet 5.5 | 15 of 15 | 15 of 15 | Claude Code agent |
| GPT-6.1 Sol | 15 of 15 | 15 of 15 | Codex agent |
| DeepSeek V4.1 Flash | 15 of 15 | 15 of 15 | API, no tools |
| Ember-1 | 15 of 15 | 15 of 15 | API, no tools |
| Gemini 3.1 Pro Preview | 15 of 15 | 15 of 15 | API, no tools |
| Muse Wonder | 15 of 15 | 15 of 15 | API, no tools |
| Kimi K3 | 15 of 15 | 15 of 15 | API, no tools |
| GPT-5.6 Luna | 15 of 15 | 15 of 15 | API, no tools |
| GPT-6 Luna | 15 of 15 | 15 of 15 | API, no tools |
| Hy4 preview | 15 of 15 | 15 of 15 | API, no tools |
| Pareto 26.10 | 15 of 15 | 15 of 15 | API, no tools |
| Claude Haiku 5.5 | 14 of 15 | 15 of 15 | API, no tools |
| MiMo-V2.6-Pro-UltraSpeed | 14 of 15 | 15 of 15 | API, no tools |
| GLM 5.2 | 14 of 15 | 15 of 15 | API, no tools |
| DeepSeek V4 Pro 0813 | 14 of 15 | 14 of 15 | API, no tools |
| GPT-5.4 | 14 of 15 | 14 of 15 | API, no tools |
| MiniMax M3 | 13 of 15 | 14 of 15 | API, no tools |
| Qwen3.7 Max | 13 of 15 | 14 of 15 | API, no tools |
| Fugu Max | 13 of 15 | 14 of 15 | API, no tools |
| Beam 1.0 501B A23B | 12 of 15 | 14 of 15 | API, no tools |
| Mistral Large 4 | 12 of 15 | 13 of 15 | API, no tools |
| Claude Haiku 4.5 | 11 of 15 | 14 of 15 | Claude Code agent |
| Nemotron 3.5 Super VL | 11 of 15 | 14 of 15 | API, no tools |
| LongCat 2.0 | 11 of 15 | 13 of 15 | API, no tools |
| Grok 4.3 | 11 of 15 | 13 of 15 | API, no tools |
| Laguna S 2.1 | 9 of 15 | 12 of 15 | API, no tools |
| Nova 2 Lite | 2 of 15 | 4 of 15 | API, no tools |

Sources: Claude Opus 5.5: [Every task, the code and Godot's output](/godot/claude-opus-5-5-godot-4-coding-test/); Claude Sonnet 5.5: [Every task, the code and Godot's output](/godot/claude-sonnet-5-5-godot-4-coding-test/); GPT-6.1 Sol: [Every task, the code and Godot's output](/godot/gpt-6-1-sol-godot-4-coding-test/); DeepSeek V4.1 Flash: [Every task, the code and Godot's output](/godot/deepseek-v4-1-flash-godot-4-coding-test/); Ember-1: [Every task, the code and Godot's output](/godot/ember-1-godot-4-coding-test/); Gemini 3.1 Pro Preview: [Every task, the code and Godot's output](/godot/gemini-3-1-pro-preview-godot-4-coding-test/); Muse Wonder: [Every task, the code and Godot's output](/godot/muse-wonder-godot-4-coding-test/); Kimi K3: [Every task, the code and Godot's output](/godot/kimi-k3-godot-4-coding-test/); GPT-5.6 Luna: [Every task, the code and Godot's output](/godot/gpt-5-6-luna-godot-4-coding-test/); GPT-6 Luna: [Every task, the code and Godot's output](/godot/gpt-6-luna-godot-4-coding-test/); Hy4 preview: [Every task, the code and Godot's output](/godot/hy4-preview-godot-4-coding-test/); Pareto 26.10: [Every task, the code and Godot's output](/godot/pareto-26-10-godot-4-coding-test/); Claude Haiku 5.5: [Every task, the code and Godot's output](/godot/claude-haiku-5-5-godot-4-coding-test/); MiMo-V2.6-Pro-UltraSpeed: [Every task, the code and Godot's output](/godot/mimo-v2-6-pro-ultraspeed-godot-4-coding-test/); GLM 5.2: [Every task, the code and Godot's output](/godot/glm-5-2-godot-4-coding-test/); DeepSeek V4 Pro 0813: [Every task, the code and Godot's output](/godot/deepseek-v4-pro-0813-godot-4-coding-test/); GPT-5.4: [Every task, the code and Godot's output](/godot/gpt-5-4-godot-4-coding-test/); MiniMax M3: [Every task, the code and Godot's output](/godot/minimax-m3-godot-4-coding-test/); Qwen3.7 Max: [Every task, the code and Godot's output](/godot/qwen3-7-max-godot-4-coding-test/); Fugu Max: [Every task, the code and Godot's output](/godot/fugu-max-godot-4-coding-test/); Beam 1.0 501B A23B: [Every task, the code and Godot's output](/godot/beam-1-0-501b-a23b-godot-4-coding-test/); Mistral Large 4: [Every task, the code and Godot's output](/godot/mistral-large-4-godot-4-coding-test/); Claude Haiku 4.5: [Every task, the code and Godot's output](/godot/claude-haiku-4-5-godot-4-coding-test/); Nemotron 3.5 Super VL: [Every task, the code and Godot's output](/godot/nemotron-3-5-super-vl-godot-4-coding-test/); LongCat 2.0: [Every task, the code and Godot's output](/godot/longcat-2-0-godot-4-coding-test/); Grok 4.3: [Every task, the code and Godot's output](/godot/grok-4-3-godot-4-coding-test/); Laguna S 2.1: [Every task, the code and Godot's output](/godot/laguna-s-2-1-godot-4-coding-test/); Nova 2 Lite: [Every task, the code and Godot's output](/godot/nova-2-lite-godot-4-coding-test/).

Read more guides: Most models now write working Godot 4 code; check the timing, and pick the tool by how much code you want to own. https://flockbay.com/blog

Related: [The best AI tools for Godot](/blog/best-ai-tools-for-godot/), [The best LLM for Godot](/blog/best-llm-for-godot/), [Godot 4 answers, verified](/godot/)