The best LLM for Godot
Eighteen of 34 AI models passed all 15 of Flockbay's checked Godot 4 coding tasks on the first try, every file run headless on the official Godot 4.7.2: Claude Opus 5.5, Claude Sonnet 5.5 and GPT-6.1 Sol as coding agents on 6 October 2026, and DeepSeek V4.1 Flash, Ember-1, Gemini 3.1 Pro Preview, Muse Wonder, Kimi K3, GPT-5.6 Luna, GPT-6 Luna, Hy4 preview and Pareto 26.10 through their APIs on 18 October, Claude Fable 5.2 through its API on 9 November, Gemini 4 Carbon through its API on 11 November, Gemini 4 Hydrogen through its API on 13 November, Qwen4 Max through its API on 16 November, and GPT-6.2 Sol and GPT-6.2 Astra through their APIs on 18 November. In game builds, Opus 5.5 finished ahead of GPT-6.1 Sol in three of five and cost 3 to 11.5 times as much at API prices: Opus for the hardest systems, Sol for everyday Godot work. Flockbay runs them on your own Claude or ChatGPT, or on its hosted AI if you have neither, free for your first build.
AppImage, x86_64. Ubuntu 22.04, Debian 12, Fedora 39 or newer.
Intel or AMD Chromebook, with Linux turned on.
Eighteen of 34 models passed all 15 Godot 4 tasks on the first try
The runs of Gemini 4 Carbon, Gemini 4 Hydrogen, Qwen4 Max, GPT-6.2 Sol and GPT-6.2 Astra are rows in the table below. Each of the other thirteen that passed all 15 on the first try has a bench page with every task, the GDScript it wrote and what Godot printed: Claude Opus 5.5, Claude Sonnet 5.5, GPT-6.1 Sol, Claude Fable 5.2, DeepSeek V4.1 Flash, Ember-1, Gemini 3.1 Pro Preview, Muse Wonder, Kimi K3, GPT-5.6 Luna, GPT-6 Luna, Hy4 preview and Pareto 26.10. Four of the 34 ran on 6 October 2026 as coding agents, allowed to read and write files and nothing else: Claude Code for the three Claude models, Codex for GPT-6.1 Sol. The other 30 ran through their APIs with no tools, 24 on 18 October, Claude Fable 5.2 on 9 November, Gemini 4 Carbon on 11 November, Gemini 4 Hydrogen on 13 November, Qwen4 Max on 16 November and GPT-6.2 Sol and GPT-6.2 Astra on 18 November, 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 fifteen models asked with no tools, and Claude Haiku 4.5 as an agent scored 11. Price did not sort them either: three of the five cheapest hosted models measured, by their usage rate in Flockbay, passed all 15, DeepSeek V4.1 Flash, GPT-6 Luna and GPT-5.6 Luna. Claude Haiku 5.5, MiMo-V2.6-Pro-UltraSpeed and GLM 5.2 passed 14, then all 15 after one fix. The spread is at the bottom, where five models failed four or more tasks on the first try and Nova 2 Lite passed two.
Each task asks for one GDScript file with a stated API, in a fresh Godot project: move a CharacterBody2D with an input action, save a dictionary as JSON, paint cells on a TileMapLayer, cast a physics ray, find a path on a grid around walls, wait for a signal or a timeout. A check scene loads the file and drives it headless on the official Godot 4.7.2 at a fixed 60 frames per second. It presses input actions, steps physics, waits on timers and signals, and compares what happened with what the task asked for. A task that fails gets the check's output once, and one chance to fix the file.
Every task touches an API that changed between Godot 3 and Godot 4, or arrived since: move_and_slide, FileAccess, signals, tweens, Timer nodes, PackedScene, TileMapLayer, ray queries, AStarGrid2D, AnimationLibrary and lambdas. The 15 tasks, their checks and the runner are open source under MIT at github.com/flockbay/godot-llm-bench, so any model can be run against the same tests.
| Model | First try | After one fix | How it ran | Run on |
|---|---|---|---|---|
| Claude Opus 5.5 | 15 of 15 | 15 of 15 | Claude Code agent | 6 Oct 2026 |
| Claude Sonnet 5.5 | 15 of 15 | 15 of 15 | Claude Code agent | 6 Oct 2026 |
| GPT-6.1 Sol | 15 of 15 | 15 of 15 | Codex agent | 6 Oct 2026 |
| Claude Fable 5.2 | 15 of 15 | 15 of 15 | API, no tools | 9 Nov 2026 |
| Gemini 4 Carbon | 15 of 15 | 15 of 15 | API, no tools | 11 Nov 2026 |
| Gemini 4 Hydrogen | 15 of 15 | 15 of 15 | API, no tools | 13 Nov 2026 |
| GPT-6.2 Astra | 15 of 15 | 15 of 15 | API, no tools | 18 Nov 2026 |
| GPT-6.2 Sol | 15 of 15 | 15 of 15 | API, no tools | 18 Nov 2026 |
| Qwen4 Max | 15 of 15 | 15 of 15 | API, no tools | 16 Nov 2026 |
| DeepSeek V4.1 Flash | 15 of 15 | 15 of 15 | API, no tools | 18 Oct 2026 |
| Ember-1 | 15 of 15 | 15 of 15 | API, no tools | 18 Oct 2026 |
| Gemini 3.1 Pro Preview | 15 of 15 | 15 of 15 | API, no tools | 18 Oct 2026 |
| Muse Wonder | 15 of 15 | 15 of 15 | API, no tools | 18 Oct 2026 |
| Kimi K3 | 15 of 15 | 15 of 15 | API, no tools | 18 Oct 2026 |
| GPT-5.6 Luna | 15 of 15 | 15 of 15 | API, no tools | 18 Oct 2026 |
| GPT-6 Luna | 15 of 15 | 15 of 15 | API, no tools | 18 Oct 2026 |
| Hy4 preview | 15 of 15 | 15 of 15 | API, no tools | 18 Oct 2026 |
| Pareto 26.10 | 15 of 15 | 15 of 15 | API, no tools | 18 Oct 2026 |
| Claude Haiku 5.5 | 14 of 15 | 15 of 15 | API, no tools | 18 Oct 2026 |
| MiMo-V2.6-Pro-UltraSpeed | 14 of 15 | 15 of 15 | API, no tools | 18 Oct 2026 |
| GLM 5.2 | 14 of 15 | 15 of 15 | API, no tools | 18 Oct 2026 |
| DeepSeek V4 Pro 0813 | 14 of 15 | 14 of 15 | API, no tools | 18 Oct 2026 |
| GPT-5.4 | 14 of 15 | 14 of 15 | API, no tools | 18 Oct 2026 |
| MiniMax M3 | 13 of 15 | 14 of 15 | API, no tools | 18 Oct 2026 |
| Qwen3.7 Max | 13 of 15 | 14 of 15 | API, no tools | 18 Oct 2026 |
| Fugu Max | 13 of 15 | 14 of 15 | API, no tools | 18 Oct 2026 |
| Beam 1.0 501B A23B | 12 of 15 | 14 of 15 | API, no tools | 18 Oct 2026 |
| Mistral Large 4 | 12 of 15 | 13 of 15 | API, no tools | 18 Oct 2026 |
| Claude Haiku 4.5 | 11 of 15 | 14 of 15 | Claude Code agent | 6 Oct 2026 |
| Nemotron 3.5 Super VL | 11 of 15 | 14 of 15 | API, no tools | 18 Oct 2026 |
| LongCat 2.0 | 11 of 15 | 13 of 15 | API, no tools | 18 Oct 2026 |
| Grok 4.3 | 11 of 15 | 13 of 15 | API, no tools | 18 Oct 2026 |
| Laguna S 2.1 | 9 of 15 | 12 of 15 | API, no tools | 18 Oct 2026 |
| Nova 2 Lite | 2 of 15 | 4 of 15 | API, no tools | 18 Oct 2026 |
The benchmark: github.com/flockbay/godot-llm-bench.
More on this: How to make games in Codex: the Codex CLI, the ChatGPT app, or your ChatGPT plan inside Flockbay
The tasks that break models are timing, not syntax
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. Across these 34 models that has mostly gone. Leaving out Nova 2 Lite, the task that failed most on the first try was waiting for a signal or a timeout, whichever comes first: 13 of the other 33 got it wrong. Five returned false for a signal 0.2 seconds away with a 1 second timeout, four never finished inside 10 seconds of game time, and four did not parse.
Building and playing an animation in code came next, with 10 failures, all the same mistake: the sprite was still visible at every sample, so it 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, twice by emitting died twice. The Godot 3 habits are now the rare failures, and they are the ones the Godot 4 answers cover: FileAccess.MODE_WRITE, which does not exist, JSON.parse() called as a static function, and a := declaration inferred from a Variant.
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.
Three models passed all 15 Godot 4 tasks as coding agents, and Haiku 4.5 missed four
Asked for a Node2D that blinks, Claude Haiku 4.5 called AnimationPlayer.add_animation, and Godot 4.7.2 stopped on the line: Nonexistent function 'add_animation' in base 'AnimationPlayer'. That call is from Godot 3. In Godot 4 an animation goes into an AnimationLibrary first. It was one of four tasks out of 15 that Haiku 4.5 got wrong on the first try. Claude Opus 5.5, Claude Sonnet 5.5 and GPT-6.1 Sol got none wrong.
Haiku 4.5's other three misses are the usual Godot 4 traps. Its health script emitted the died signal twice. Its grid path returned an untyped Array from a function declared Array[Vector2i], and Godot stopped it with a script error. Its coroutine that waits for a signal or a timeout never finished inside 10 seconds of game time. Shown the error, it fixed three of the four. The coroutine still failed: it returned false for a signal 0.2 seconds away with a 1 second timeout.
Speed was close at the top. The median task took the agent 8 seconds on Opus 5.5 and Sonnet 5.5, 10 on GPT-6.1 Sol and 14 on Haiku 4.5, timed from the agent's start to its reply, so Claude Code's and Codex's own start-up are inside it. Claude Haiku 4.5's bench page has each of its four misses with what Godot printed.
More on this: Best AI game maker in 2026
In game builds, Opus 5.5 finished ahead more often and GPT-6.1 Sol cost a fraction
Fifteen small scripts say whether a model knows Godot 4. They do not say whether it can build a game system and test it. So the same models took on bigger jobs in Flockbay over MCP: a physics truck, a six-legged walker, doors, a rigged character twice and a save system, with the same prompt for every model in a build. They pressed Play themselves and decided when they were done. Then both sides were played or filmed the same way, and each video ended on a verdict.
Opus 5.5 took the physics truck, with a trail more than twice as long, and the doors, where one rule about what counts as in the way jammed Sol's doors and stranded its guard. It also came first of four on a rigged character built with Flockbay's modeler. GPT-6.1 Sol took the same character built in Blender, and beat Sonnet 5.5 on a six-legged walker: Sonnet's could walk into a spot it never walked out of. Save and load was a draw, with both games matching the game left running on 25 of 26 loads.
The bill went the other way. Across the five builds Opus and Sol both played, Opus's runs came to $2.93 to $10.48 at public API prices and Sol's to $0.63 to $1.55, so Opus cost 3 to 11.5 times as much for the same job. Sonnet 5.5's walker cost about seven times Sol's at the same price per token. Every one of these runs went through a subscription, so the dollar figures are estimates at API prices, not what the runs were billed.
Claude Opus 5.5. Verdict: the strongest Godot model in these game builds, 15 of 15 and ahead in three of five, at 3 to 11.5 times the cost of GPT-6.1 Sol.
GPT-6.1 Sol. Verdict: the value pick, 15 of 15 with two build wins and a draw, each of its builds against Opus costing $0.63 to $1.55 at API prices.
Claude Sonnet 5.5. Verdict: level with Opus on short GDScript, 15 of 15 at the same 8-second median, but it lost the six-legged walker to Sol at about seven times the cost.
Claude Haiku 4.5. Verdict: fine for plain scripts when it sees Godot's errors, at 11 of 15 on the first try and 14 after one fix.
A GDScript edit, a scene edit and an hour-long agent loop want different models
A GDScript function is a short coding task with a strict runtime. The model has to emit Godot 4 names, typed where the project is typed, and a signal connection that actually exists. On that job sixteen models tie at 15 of 15, so pick on speed and price. Of the four agents, Opus 5.5 and Sonnet 5.5 had the lowest median time, 8 seconds, and on 3 October Anthropic's pages listed Sonnet 5.5 at half the API price of Opus. Haiku 4.5 handles the plain ones, a mover, a label, a save file. Check its signals, typed arrays and coroutines, because three of its four misses were there, and the fourth was a Godot 3 call.
Scene editing through MCP is tool use. The model reads a scene, calls a tool, reads the tool result and calls the next one. Every build above was that job, done through Flockbay's MCP tools, and on the doors the difference was how hard each model tested its own work. Opus made 67 model calls and 5 Plays and wrote a three-minute stress test with the player in the doorways. Sol made 23 calls and 2 Plays, and its 46 checks moved the player out of the house first. The Claude docs read on 3 October 2026 say Opus 5.5 and Sonnet 5.5 return an error when a request forces a specific tool, so an agent loop on them should let the model choose.
A long agent loop is a context-window problem before it is a quality problem. Every tool result stays in the conversation until something summarises it. On their own pages, read on 3 October 2026, Claude Opus 5.5, Claude Sonnet 5.5 and Gemini 3.8 Flash each publish a window of about a million tokens, and OpenAI publishes 1,050,000 for GPT-5.5 on the API but 400K for GPT-5.5 in Codex. The builds show the cost side of the same loop. On the truck, the walker and the doors, Opus and Sonnet re-read 8 to 17 million cached tokens a run, and Sol 1.3 to 2.9 million. On a subscription that is usage. On an API key it is the bill.
More on this: Best AI game maker
A cheap model writes working Godot 4 code when it sees the error
Haiku 4.5 went from 11 of 15 to 14 of 15 with one look at Godot's output. That is the useful finding for a cheap model: the error loop matters more than the first draft. A model that runs the game, reads the GDScript parse errors and the debugger, and tries again recovers most of what it got wrong.
Claude Haiku 5.5 came out after the agent run and took the 15 tasks through its API on 18 October: 14 on the first try, missing the animation built in code, and all 15 after one fix. In its build log it built a Celeste-style jump in Godot over Flockbay's MCP server, with the measurements sent back after each pass and never the fix. Its first pass met 4 of 6 bars, coyote time and jump buffering among them. Its third met all 6. The three passes came to about $0.86 at API prices, and the same tokens on Opus 5.5 would have come to about $7.60.
A local model is the cheapest of all, and none ran these tasks. If you run one, give it the same loop: a headless run and the full error text after every change. Making games with local LLMs covers the setup.
One evening on your own game beats any comparison, this one included
Pick one task 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.
A small project tells you more in an evening than a table can. Every model's bench page has the code it wrote and Godot's output, so you can see exactly where it went wrong before you trust it with yours.
The model is half the answer, and the tool around it decides how much code you own
Asked as the best AI for Godot, the question is also which kind of tool puts the model to work. The best AI tools for Godot has each one with a price and a verdict. The kinds come down to how much of the code you want to own.
A coding assistant such as Copilot, Cursor or Claude Code, pointed at your project folder, leaves you the programmer: it drafts and fixes, and it does best when you name your Godot version and paste errors in full. An editor plugin sees the scene and script you have open. 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, and still leaves you to judge what it built.
Flockbay is a free desktop app for Windows, macOS and Linux that works with Godot 4.6+, with the official Godot inside. Its AI writes the GDScript, runs the game to check its work, reads Godot's errors 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.
A ranking is true on the day it was run, so this one is rerun
A table that names one best model for Godot is wrong the day the next model ships. Claude Opus 5.5 came out on 22 September 2026 and Claude Sonnet 5.5 on 28 September. Six days apart, both replaced the recommendation a spring article would still be printing.
So every score here carries the day it was measured. The four coding agents ran the 15 tasks on 6 October 2026, 24 models ran through their APIs on 18 October, Claude Fable 5.2 on 9 November, Gemini 4 Carbon on 11 November, Gemini 4 Hydrogen on 13 November, Qwen4 Max on 16 November and GPT-6.2 Sol and GPT-6.2 Astra on 18 November, and the game builds went out between 5 and 10 October. When a model ships, the same tasks run again on the engine Flockbay ships, the model gets its own dated row, and this URL stays the same.
What lasts longer than a score is the job. A GDScript function, a scene edit through an MCP tool call and an agent loop that runs for an hour do not want the same thing from a model. The vendor sections below are dated to the page they were read from.
Where these numbers stop
The 15 tasks are small, single-file and checked by running them. They say whether a model's Godot 4 code works. They do not say whether it designs a good level, picks a good art direction or keeps a 50-file project coherent.
Each game build is one run per model and one reviewer's verdict, so treat a single win as a data point, not a ranking. The times are the four agents', start-up included, so Claude Code and Codex are part of what they measure. The 24 models asked through their APIs had no tools and one request per attempt, so their score is the model's own GDScript with no agent around it. And the dollar figures are estimates at public API prices for runs that went through subscriptions.
Claude Opus 5.5 is built for long agentic runs, at $4 and $20 per million tokens
Dated 3 October 2026, from Anthropic's model page. Released 22 September 2026. Model id claude-opus-5-5. Anthropic describes it as built for long-running agentic coding and knowledge work. Context window 1M tokens, max output 128K, adaptive thinking always on, default effort medium, knowledge cutoff June 2026. API price on that page: $4 per million input tokens and $20 per million output tokens.
For a long Godot session that is the relevant line: long-running agentic coding, and a 1M window. The limitation on the same page is that forced tool use returns an error, and thinking cannot be turned off. The page does not mention Godot, GDScript or MCP. Flockbay's runs are the Godot evidence: 15 of 15, and ahead of GPT-6.1 Sol in three of five game builds.
Claude Sonnet 5.5 has the same window at half the price of Opus
Dated 3 October 2026, from Anthropic's model page. Released 28 September 2026. Model id claude-sonnet-5-5. Anthropic calls it the best combination of speed and intelligence. Context window 1M tokens, max output 128K, adaptive thinking, default effort high, comparative latency fast, knowledge cutoff June 2026. API price on that page: $2 per million input tokens and $10 per million output tokens.
Forced tool use returns an error on this model too. Setting temperature, top_p or top_k off the default returns a 400. The page does not mention Godot. In Flockbay's runs it passed 15 of 15 at the lowest median time, then lost the six-legged walker to GPT-6.1 Sol at about seven times Sol's cost.
More on this: Godot MCP: connect Claude Code, Codex or Cursor to your Godot game
GPT-5.5 has a million tokens on the API and 400K in Codex
Dated 3 October 2026, from OpenAI's API model page and the 23 April 2026 announcement. The API page calls GPT-5.5 a flagship model for complex professional work, with a 1,050,000-token context window, 128,000 max output tokens, and a knowledge cutoff of 1 December 2025. API price on that page: $5 per million input tokens and $30 per million output tokens. Prompts over 272K input tokens are priced higher for the whole session.
The announcement says GPT-5.5 in Codex has a 400K context window, on Plus, Pro, Business, Enterprise, Edu and Go. A ChatGPT sign-in runs in Flockbay on any plan, Free included, and Codex is how it builds. Neither page publishes a Godot score. GPT-5.5 itself was not in Flockbay's runs: GPT-6.1 Sol ran in Codex, and GPT-6 Luna, GPT-5.6 Luna and GPT-5.4 through the API. OpenAI's page for GPT-6.1 Sol was not read for this edition.
Gemini 3.8 Flash publishes a million-token window and no Godot score
Dated 3 October 2026, from Google's Gemini API model page, which lists Gemini 3.8 Flash as stable. Google describes it as its most intelligent Flash model, for long-horizon software engineering and autonomous agents. Input token limit 1,048,576. Output token limit 65,536. Function calling, code execution and structured outputs are supported. Computer use is marked preview. Thinking supports low, medium and high, and the page says minimal returns an error. The page's latest-update line says September 2026. It does not mention Godot.
Gemini is not one of the sign-ins in Flockbay's chat, and Gemini 3.8 Flash was not in Flockbay's runs. Gemini 3.1 Pro Preview was, through its API on 18 October, and passed all 15 tasks on the first try. Gemini CLI connects over Flockbay's MCP server like any other agent. The Flash model belongs in this comparison because the search is for a model.
Qwen3-Coder is the open-weight pick, with the smallest window here
Dated from the Qwen team's post of 22 July 2025, which is the open-weight card this edition uses. Qwen3-Coder-480B-A35B-Instruct is a mixture-of-experts model, 480B parameters with 35B active. The post says it supports 256K tokens of context natively and 1M with extrapolation, and that it was trained for agentic coding and tool use. The same post compares it with Claude Sonnet 4, which is two generations behind the Claude models above. Treat that comparison as historical.
256K native context is the smallest window in this set. A long Godot agent loop that keeps tool results will fill it before it fills a 1M window. Running the weights yourself is the reason to pick it: the machine is yours, and the vendor page is not metering the turn. Qwen3-Coder was not in Flockbay's runs, and no Godot score is on the post. Qwen3.7 Max was, through its API on 18 October: 13 of 15 on the first try and 14 after one fix.
Claude, ChatGPT and your own agent run in Flockbay, free to start
Flockbay builds and play-tests a whole Godot game, free for your first build on Flockbay's AI or your own Claude or ChatGPT, on the official Godot engine. On a Claude plan, Pro or Max, the model menu lists the models Claude Code offers your account. Signed in with ChatGPT, on any plan, Free included, it lists your account's models. Your own agent, such as Claude Code, Codex or Cursor, connects over Flockbay's MCP server and brings its own model. With none of these, Flockbay's hosted model, Flockbay in the model menu, is free for your first build. After it, plans add hosted AI, and new accounts build on their own AI on a plan. Claude Free does not run here. The setup is Claude or ChatGPT with Godot.
Four models on the same Godot 4 work, measured by Flockbay
Flockbay's own runs, not vendor claims. The 15 coding tasks ran headless on the official Godot 4.7.2 at a fixed 60 frames per second on 6 October 2026, with one fix allowed after the test's output, and the time is the coding agent's, start-up included. The head-to-heads are game builds in Flockbay over MCP, published between 5 and 10 October 2026. Costs are estimates at public API prices for runs that went through a subscription.
| Model | Godot 4 tasks passed, first try | After one fix | Median time per task | Agent | Head-to-head builds | Estimated API cost per build | Read from |
|---|---|---|---|---|---|---|---|
| Claude Opus 5.5 | 15 of 15 | 15 of 15 | 8 s | Claude Code | Took the physics truck and the doors from GPT-6.1 Sol, and came first of four on the modeled character. Lost the Blender character to Sol. Drew save and load. | $2.93 to $10.48 over five builds | Coding test · Truck · Doors · Four models · Blender character · Save and load |
| GPT-6.1 Sol | 15 of 15 | 15 of 15 | 10 s | Codex | Took the Blender character from Opus 5.5 and the six-legged walker from Sonnet 5.5. Lost the truck and the doors to Opus, and finished behind it on the modeled character. Drew save and load. | $0.63 to $1.55 over the same five builds; $0.86 for the walker | Coding test · Truck · Walker · Doors · Four models · Blender character · Save and load |
| Claude Sonnet 5.5 | 15 of 15 | 15 of 15 | 8 s | Claude Code | Lost the six-legged walker to GPT-6.1 Sol: its walker could walk into a spot it never walked out of. On the modeled character, no hands and a pickaxe through the floor every other step. | $6.05 for the walker's first pass; $0.64 for the character | Coding test · Walker · Four models |
| Claude Haiku 4.5 | 11 of 15 | 14 of 15 | 14 s | Claude Code | In its one head-to-head, the modeled character, it made a figure of stacked blocks. | $0.45 for the character | Coding test · Four models |
Questions
- What is the best AI model for Godot?
- On Flockbay's 15 checked Godot 4 coding tasks, run on the official Godot 4.7.2, eighteen of 34 models passed all 15 on the first try: Claude Opus 5.5, Claude Sonnet 5.5 and GPT-6.1 Sol as coding agents on 6 October 2026, and DeepSeek V4.1 Flash, Ember-1, Gemini 3.1 Pro Preview, Muse Wonder, Kimi K3, GPT-5.6 Luna, GPT-6 Luna, Hy4 preview and Pareto 26.10 through their APIs on 18 October, Claude Fable 5.2 through its API on 9 November, Gemini 4 Carbon through its API on 11 November, Gemini 4 Hydrogen through its API on 13 November, Qwen4 Max through its API on 16 November, and GPT-6.2 Sol and GPT-6.2 Astra through their APIs on 18 November. In game builds, Opus 5.5 finished ahead of GPT-6.1 Sol in three of five, and its runs cost 3 to 11.5 times as much at API prices. Opus for the hardest systems, Sol for everyday work.
- Which AI is best at GDScript?
- On short GDScript tasks sixteen models tie at 15 of 15 on the first try, among them Opus 5.5, Sonnet 5.5 and GPT-6.1 Sol, whose median agent times were 8, 8 and 10 seconds. Where models missed, it was mostly timing: a coroutine racing a signal against a timer, an animation track that never blinked, a signal emitted twice. Godot 3 calls such as FileAccess.MODE_WRITE are now the rare failures.
- Which Godot 4 tasks do AI models get wrong most?
- Timing. Leaving out one outlier, 13 of 33 models failed a coroutine that waits for a signal or a timeout on the first try, 10 built an animation in code that never blinked, and 6 failed a path on a grid around walls. Health with signals failed 3 times. Saving a dictionary as JSON failed 3 times, on FileAccess.MODE_WRITE, on JSON.parse() called as a static function and on a type inferred from a Variant.
- Is Claude or ChatGPT better for Godot?
- Neither wins everywhere. Claude Opus 5.5 beat GPT-6.1 Sol on a physics truck, on doors and on a character built with Flockbay's modeler. Sol won the same character built in Blender and beat Claude Sonnet 5.5 on a six-legged walker, and save and load was a draw. Against Opus, Sol was the cheaper side every time: Opus cost 3 to 11.5 times as much at API prices.
- Can a cheap model write Godot 4 code?
- Mostly, if it sees Godot's errors. Three of the five cheapest hosted models measured, DeepSeek V4.1 Flash, GPT-6 Luna and GPT-5.6 Luna, passed all 15 tasks on the first try. Claude Haiku 4.5 passed 11 and then 14 after one fix with the test's output, and Claude Haiku 5.5 passed 14 and then all 15. In a build log, Haiku 5.5 built a Celeste-style jump that met all six of its bars by the third pass, for about $0.86 at API prices.
- How often is this measured?
- The four coding agents ran the 15 tasks on 6 October 2026, 24 models ran through their APIs on 18 October, Claude Fable 5.2 on 9 November, Gemini 4 Carbon on 11 November, Gemini 4 Hydrogen on 13 November, Qwen4 Max on 16 November and GPT-6.2 Sol and GPT-6.2 Astra on 18 November, and the game builds went out between 5 and 10 October. The tasks run again when a model ships, each score keeps the day it was measured, and this URL stays the same.
- Which plan runs in Flockbay?
- Flockbay's hosted model, free for your first build. Also Claude Pro or Max, any ChatGPT plan, Free included, your own agent over MCP, or your own API key, free through your first build. Claude Free does not run here. Flockbay's own plans, Hobby, Indie Developer and Studio, add more hosted AI and, for new accounts, your own AI after the first build.
- Does a bigger context window win a long agent loop?
- It wins the part of the loop that is "did the earlier tool result survive". Opus 5.5, Sonnet 5.5 and Gemini 3.8 Flash publish about a million tokens. GPT-5.5 publishes 1,050,000 on the API and 400K in Codex. Qwen3-Coder publishes 256K native.
Flockbay's AI or your own Claude or ChatGPT, free for your first build, on the official Godot engine. Windows, macOS and Linux.
Use Claude or ChatGPT with Godot, free to startAppImage, x86_64. Ubuntu 22.04, Debian 12, Fedora 39 or newer.
Intel or AMD Chromebook, with Linux turned on.
