For terminal work, choose GPT-6 Astra in Codex or Claude Fable 5 in Claude Code: these model-agent pairs score 68.8% and 67.3% on 333 GameDevBench Godot 4 tasks . For an assistant inside Godot, consider Ziva, which we make, or AI Assistant Hub. Prices and versions checked September 19, 2026.
In the self-selected Godot Community Poll 2026, 63.2% of respondents mainly use Godot’s built-in code editor. Cursor, Devin Desktop (formerly Windsurf), Copilot’s official clients and JetBrains AI work outside that editor.
For more plugins, see Best AI Tools for Godot in 2026. The scores come from GameDevBench, and the features and prices come from product documentation. We checked sample Godot 3 scripts with Godot 4.7.2’s command-line parser on Linux. We did not test assistants, local setups, project instructions or the prompts below.
TL;DR #
| If you… | Pick |
|---|---|
| code in Godot’s built-in editor | Ziva (free Hobby plan ),AI Assistant Hub (free, MIT) , or a terminal agent beside Godot |
| want autocomplete in a separate editor | Copilot Free ,Devin Desktop Free orCursor Pro |
| want multi-file agents | GPT-6 Astra in Codex or Claude Fable 5 in Claude Code; add a Godot MCP server, a bridge that lets agents use editor tools |
| write C# in Godot | Rider (free for non-commercial use) withCopilot (needs a Copilot subscription) or JetBrains AI |
| want to run local models | AI Assistant Hub, Ziva orClaude Code through Ollama | Ziva requires internet for AI features; its local-model docs do not say whether those models work offline.
For autocomplete inside Godot, a community Copilot plugin requires Node.js 20.8+ and a Copilot subscription, including Free. The official godot-tools extension adds GDScript support to VS Code and is also on Open VSX , an extension registry. Cursor installs extensions from Open VSX .
What GameDevBench shows #
GameDevBench , by Carnegie Mellon and Princeton researchers at ICML 2026, derives its tasks from Godot 4 tutorials. Its leaderboard mixes models, agent software and submission dates, so the scores cannot isolate the assistant’s contribution.
These are the highest listed entries for each agent. Missing ranks are other entries from agents already listed, mostly Codex and Claude Code. The score is pass@1: the percentage of tasks solved on the first attempt. The 95% confidence interval shows the uncertainty around that score; high and xhigh are reasoning effort settings.
| Rank | Agent | Model entry | pass@1, 95% CI |
|---|---|---|---|
| 1 | Codex | gpt-6-astra (high) | 68.8% ±5.0 |
| 2 | Claude Code | claude-fable-5 (xhigh) | 67.3% ±5.0 |
| 5 | Muse Code | muse-spark-1.2 [high] | 61.0% ±5.2 |
| 7 | Kimi Code | kimi-k3 | 58.0% ±5.3 | | 10 | Gemini CLI | gemini-3-pro-preview | 53.8% ±5.4 | | 14 | OpenCode | glm-5.2 | 38.4% ±5.2 | | 16 | OpenHands | kimi-k2.5 | 20.7% ±4.4 |
The overlapping confidence intervals do not establish a clear winner among Codex, Claude Code and Muse Code. Each entry uses its model’s best agent software and feedback setup, which can include editor screenshots and gameplay video. OpenHands reached 38.4% with GPT-5.4 Mini without visual feedback, above its 20.7% leaderboard entry with Kimi K2.5.
Even the leader failed 104 of 333 tasks. Common failures involve scene structure, signals (messages between objects) and resources (data objects used by nodes and scripts). Treat the rankings as a shortlist, then check the assistant against your own project.
The benchmark excludes Copilot, Cursor, Devin Desktop, JetBrains AI and in-editor plugins, including Ziva and AI Assistant Hub. It uses Godot 4.4.1 , while 63.8% of the community poll’s respondents mainly use 4.7.
Disclosure: GPT-6 Astra wrote this post from a Claude draft, with research and fact-checking by Claude agents. GPT-6 Astra and Claude Fable 5 rank first and second.
Can the assistant see your editor and game? #
The paper’s Table 2 compares visual feedback across eleven model-agent pairs. Eight scores rose with editor screenshots plus gameplay video; three fell. Most changes are small compared with the confidence intervals. GPT-5.4 showed a larger gain, from 41.1% to 52.0%.
Researchers supplied screenshots through their own MCP server and instructions for recording gameplay video. The leaderboard instead keeps each model’s best feedback setup.
| Agent | Model | Without visual feedback | Screenshots + video |
|---|---|---|---|
| Codex | GPT-5.4 | 41.1% | 52.0% |
| Claude Code | Sonnet 4.5 | 28.8% | 34.8% |
| Gemini CLI | Gemini 3 Pro Preview | 50.1% | 53.8% |
| Claude Code | Haiku 4.5 | 13.8% | 16.5% |
| Codex | GPT-5.4 Mini | 36.9% | 39.0% |
| OpenHands | Kimi K2.5 | 18.9% | 20.7% |
| OpenHands | Qwen 3.5 397B A17B | 5.4% | 5.1% |
| Gemini CLI | Gemini 3 Flash Preview | 45.4% | 44.1% |
| OpenHands | Haiku 4.5 | 15.6% | 17.7% |
| OpenHands | GPT-5.4 Mini | 38.4% | 36.9% |
| OpenHands | Gemini 3 Flash Preview | 30.3% | 31.8% |
The paper tested its own server, not these public bridges. Per its docs, Coding-Solo/godot-mcp runs projects and captures debug output. The third-party open-source MCP plugin hi-godot/godot-ai connects Claude Code and Codex to a live editor to edit scenes, nodes and scripts. Ziva’s MCP server lets external agents inspect the scene tree (the hierarchy of nodes), read errors and run games.
AI Assistant Hub enables agent tools only with Ollama or llama.cpp. Its tools scan scenes and manage nodes .
Isaac Dedini built his card-game UI entirely through Claude Code, opening the Godot editor only once. His custom test runner let Claude compare screenshots against the intended UI.
Avoiding Godot 3 code in Godot 4 #
Developer reports vary. In March, Ariarule reported good GDScript from Opus 4.5, 4.6 and Sonnet 4.6 using CLAUDE.md. Ariarule still saw occasional Godot 3 output. In September, ieishi reported mixed syntax despite requesting Godot 4, without naming the tool.
Put your engine version and project conventions in a root AGENTS.md. Codex, Cursor and Ziva read it.
For Claude Code, add a line containing @AGENTS.md to CLAUDE.md. This imports your shared instructions; see the memory docs for automatic rules. Gemini CLI uses GEMINI.md.
Copilot chat in VS Code reads AGENTS.md, but inline suggestions ignore it.
Include rules specific to your game. For example, MrPhil’s Stellar Throne instructions prohibit await in manager _ready() functions to prevent load-order bugs.
Godot 4 changed classes, signals, Tween and the tool keyword . It replaced yield, export and onready with await and annotations . Use this as a reference when writing your AGENTS.md rules:
| Godot 3 | Godot 4 |
|---|---|
KinematicBody2D |
CharacterBody2D |
KinematicBody |
CharacterBody3D |
Spatial |
Node3D |
yield(...) |
await |
export var |
@export var |
onready var |
@onready var |
| `instance()` | `instantiate()` |
| Tween node | [`create_tween()`](https://docs.godotengine.org/en/stable/classes/class_node.html) |
| `connect("sig", obj, "method")` | `sig.connect(callable)` |
| tool | @tool |
A five-minute test #
- Ask for a player controller matching your engine version and naming conventions. Check for the old forms on the left above.
- Use Godot’s parse-only check :
godot --headless --check-only --script <file>.gd.godotstands for your Godot executable path. Our check rejectedyield,export var,onready var,extends KinematicBody2D, three-argumentconnect()and baretool. It passed.instance()on an untyped value, so run the game too. - Ask which scene is open and which nodes it contains. Compare with the editor to check whether the assistant knows its live state.
- Introduce an error and ask for a fix without pasting it. This checks whether errors reach the assistant automatically. Then run the game and inspect the result yourself.
Price, local models and privacy #
| Assistant | Runs in | Tool entry price | Free tier |
|---|---|---|---|
| GitHub Copilot | VS Code, Rider | Pro $10/mo | 2,000 completions a month |
| Cursor | Separate editor | Pro $20/mo | Hobby : limited Agent requests |
| Devin Desktop , formerly Windsurf | Separate editor | Pro $20/mo | Unlimited autocomplete |
| JetBrains AI | Rider: GDScript and C# | AI Pro $10/mo | 3 AI Credits per 30 days (about $3 of usage ) |
| Claude Code | Terminal, IDEs | Claude Pro $20/mo; Fable costs below | Not in Claude Free |
| Codex | Terminal, IDEs | ChatGPT Go $8/mo, Plus $20/mo; model limits below | In the $0 ChatGPT plan; model limits below |
| Gemini CLI | Terminal | Paid API key | Free login ended June 18, 2026 |
| Ziva | Godot | Basic $20/mo; hosted Astra and Fable 5.1 need Pro $50/mo | Hobby: $3 of AI usage a month |
| AI Assistant Hub | Godot | Plugin free, MIT | Free plugin; model costs depend on provider |
Codex’s pricing page gives model usage figures for Plus and higher, but none for Free or Go.
Claude’s pricing page limits Pro’s Fable access to usage credits at standard API rates. Fable 5 and 5.1: $10/million input tokens, $50/million output tokens. Max includes Fable from $100/mo, capped at half the weekly limits.
Ziva’s hosted model list puts GPT-6 Astra and Claude Fable 5.1 on Ziva’s Pro plan, $50/mo. Fable 5.1 is newer than the benchmarked Fable 5. Your own Claude Code or ChatGPT Codex subscription works on any Ziva plan and counts against that provider’s usage.
Gemini CLI’s README still lists free login; its successor Antigravity has a $0 individual plan .
Ziva and JetBrains AI support Ollama and LM Studio. JetBrains defaults to a 64,000-token context window, the text a model can consider at once. It cannot call configured MCP tools with local models. For Claude Code, run ollama launch claude and set a 64k+ context window, per the Ollama guide above.
The paper identifies Qwen3.5-397B and Kimi K2.5 as open-weight models, with downloadable model weights. They scored 5.4% and 20.7%; GLM-5.2 scored 38.4% and Kimi K3 58.0%, below the leader’s 68.8%. These benchmark scores do not come from local Ollama setups.
Check training settings before sending unreleased code:
- Copilot Free, Pro and Pro+ interaction data trains models unless users opt out .
- On free non-commercial licences, JetBrains AI collects detailed code data by default and uses it to train models. Opt out under Settings > Data Sharing, per the JetBrains AI FAQ linked above.
- For Claude Free, Pro and Max, enabling model improvement permits training and five-year retention. Disabling it means 30-day retention; the cited policy does not state a default.
- Cursor’s privacy mode prevents training on code data. Its pricing FAQ gives no default; enable it in settings.
- ChatGPT individual plans include Codex; content may train models unless users opt out .
- Ziva’s privacy policy says it stores no AI-request prompts, code or project data and does not train on code. Provider retention terms still apply.
See AI coding tools and code privacy for more on those choices.
Where Ziva fits #
Ziva writes GDScript and C# and sees scenes, editor errors and the running game. It supports Godot 4.2 or later, per its installation docs.
Godot errors sent to Ziva’s agent include file and line (plugin version 3.1.2), including errors its operations caused (3.2.5). The Playtest Agent plays games and sends input, and receives runtime errors (3.1.5).
The Playtest Agent needs Pro or Ultra, $50/mo and up.
Ziva has no inline autocomplete; for autocomplete, consider Copilot or Cursor. It is not open source: the terms prohibit extracting source code.
Related reading #
- [Is there an AI copilot inside the Godot editor?](https://ziva.sh/blogs/ai-copilot-inside-godot-editor) covers in-editor assistants
- [How do I generate GDScript with AI?](https://ziva.sh/blogs/generate-gdscript-with-ai) explains the three routes, step by step
- [Best AI Tools for Godot in 2026: 11 Plugins Compared](https://ziva.sh/blogs/best-ai-tools-for-godot-2026) covers plugins and MCP bridges
- [GDScript vs C# in Godot: Which Should You Pick?](https://ziva.sh/blogs/gdscript-vs-csharp) helps you pick the language first