Can MiniMax M3 write working Godot 4 code? 13 of 15 on the first try
By Eduardo Orellana · Updated
MiniMax M3 passed 13 of 15 Godot 4 tasks on the first try and 14 of 15 after one fix with the test's output, each run headless on Godot 4.7.2. It missed: build and play an animation in code, wait for a signal or a timeout.
AppImage, x86_64. Ubuntu 22.04, Debian 12, Fedora 39 or newer.
Intel or AMD Chromebook, with Linux turned on.
Every model measured on the same Godot 4 tasks, each run headless on Godot 4.7.2 at a fixed 60 frames per second.
Measured by Flockbay: the same tasks, the same checks, one fix allowed with the test output.
| Model | First try | After one fix | Agent | Run on | Read from |
|---|---|---|---|---|---|
| Claude Opus 5.5 | 15 of 15 | 15 of 15 | Claude Code | 2026-10-06 | Claude Opus 5.5 results |
| Claude Sonnet 5.5 | 15 of 15 | 15 of 15 | Claude Code | 2026-10-06 | Claude Sonnet 5.5 results |
| GPT-6.1 Sol | 15 of 15 | 15 of 15 | Codex | 2026-10-06 | GPT-6.1 Sol results |
| DeepSeek V4.1 Flash | 15 of 15 | 15 of 15 | Hosted, by API | 2026-10-18 | DeepSeek V4.1 Flash results |
| Ember-1 | 15 of 15 | 15 of 15 | Hosted, by API | 2026-10-18 | Ember-1 results |
| Gemini 3.1 Pro Preview | 15 of 15 | 15 of 15 | Hosted, by API | 2026-10-18 | Gemini 3.1 Pro Preview results |
| Muse Wonder | 15 of 15 | 15 of 15 | Hosted, by API | 2026-10-18 | Muse Wonder results |
| Kimi K3 | 15 of 15 | 15 of 15 | Hosted, by API | 2026-10-18 | Kimi K3 results |
| GPT-5.6 Luna | 15 of 15 | 15 of 15 | Hosted, by API | 2026-10-18 | GPT-5.6 Luna results |
| GPT-6 Luna | 15 of 15 | 15 of 15 | Hosted, by API | 2026-10-18 | GPT-6 Luna results |
| Hy4 preview | 15 of 15 | 15 of 15 | Hosted, by API | 2026-10-18 | Hy4 preview results |
| Pareto 26.10 | 15 of 15 | 15 of 15 | Hosted, by API | 2026-10-18 | Pareto 26.10 results |
| Claude Haiku 5.5 | 14 of 15 | 15 of 15 | Hosted, by API | 2026-10-18 | Claude Haiku 5.5 results |
| MiMo-V2.6-Pro-UltraSpeed | 14 of 15 | 15 of 15 | Hosted, by API | 2026-10-18 | MiMo-V2.6-Pro-UltraSpeed results |
| GLM 5.2 | 14 of 15 | 15 of 15 | Hosted, by API | 2026-10-18 | GLM 5.2 results |
| DeepSeek V4 Pro 0813 | 14 of 15 | 14 of 15 | Hosted, by API | 2026-10-18 | DeepSeek V4 Pro 0813 results |
| GPT-5.4 | 14 of 15 | 14 of 15 | Hosted, by API | 2026-10-18 | GPT-5.4 results |
| MiniMax M3 | 13 of 15 | 14 of 15 | Hosted, by API | 2026-10-18 | MiniMax M3 results |
| Qwen3.7 Max | 13 of 15 | 14 of 15 | Hosted, by API | 2026-10-18 | Qwen3.7 Max results |
| Fugu Max | 13 of 15 | 14 of 15 | Hosted, by API | 2026-10-18 | Fugu Max results |
| Beam 1.0 501B A23B | 12 of 15 | 14 of 15 | Hosted, by API | 2026-10-18 | Beam 1.0 501B A23B results |
| Mistral Large 4 | 12 of 15 | 13 of 15 | Hosted, by API | 2026-10-18 | Mistral Large 4 results |
| Claude Haiku 4.5 | 11 of 15 | 14 of 15 | Claude Code | 2026-10-06 | Claude Haiku 4.5 results |
| Nemotron 3.5 Super VL | 11 of 15 | 14 of 15 | Hosted, by API | 2026-10-18 | Nemotron 3.5 Super VL results |
| LongCat 2.0 | 11 of 15 | 13 of 15 | Hosted, by API | 2026-10-18 | LongCat 2.0 results |
| Grok 4.3 | 11 of 15 | 13 of 15 | Hosted, by API | 2026-10-18 | Grok 4.3 results |
| Laguna S 2.1 | 9 of 15 | 12 of 15 | Hosted, by API | 2026-10-18 | Laguna S 2.1 results |
| Nova 2 Lite | 2 of 15 | 4 of 15 | Hosted, by API | 2026-10-18 | Nova 2 Lite results |
How the test works
Each task asks for one GDScript file with a stated API, in a fresh Godot project. MiniMax M3 was asked through its API, the way Flockbay reaches its hosted models: the task in one request, no tools, and the whole file back in a code block. A check scene then loaded the file and drove it on Godot 4.7.2, headless, at a fixed 60 frames per second: it pressed input actions, stepped physics, waited on timers and signals, and compared the results.
When a task failed, the model got the check's output once and could fix its file. Every task touches an API that changed between Godot 3 and Godot 4, or was added since: move_and_slide and is_on_floor, FileAccess and JSON, signals, tweens, Timer nodes, PackedScene, TileMapLayer, ray queries, AStarGrid2D, AnimationLibrary, and lambdas, which capture local variables by value.
The checks themselves are tested first: each passes on a reference answer and fails on an empty script. Run on 2026-10-18.
MiniMax M3 in Flockbay
MiniMax M3 is a hosted model in the Flockbay app's model menu, on the Hobby plan and above. It uses the plan's hosted usage about 4.6 times as fast as Auto.
It takes images, so Flockbay can show it screenshots of your game, and it has an effort setting in the composer, from quick answers to longer reasoning.
Move a CharacterBody2D with an input action: pass
The task: Write res://player.gd for Godot 4: a script for a CharacterBody2D that moves right at 200 pixels per second while the input action "ui_right" is held, and stands still otherwise.
Passed on the first try.
extends CharacterBody2D
const SPEED = 200.0
func _physics_process(_delta):
if Input.is_action_pressed("ui_right"):
velocity.x = SPEED
else:
velocity.x = 0.0
move_and_slide()
Fall with gravity and land on the floor: pass
The task: Write res://ball.gd for Godot 4: a script for a CharacterBody2D that falls with a gravity of 980 pixels per second squared and stops when it lands on the floor. The scene gives it its collision shape.
Passed on the first try.
extends CharacterBody2D
const GRAVITY := 980.0
func _physics_process(delta: float) -> void:
if is_on_floor():
velocity.y = 0.0
else:
velocity.y += GRAVITY * delta
move_and_slide()
A score label: pass
The task: Write res://score_label.gd for Godot 4: a script for a Label that shows "Score: 0" when it enters the scene, and has a method add(points: int) that adds to the score and updates the text to "Score: <total>".
Passed on the first try.
extends Label
var score: int = 0
func _ready() -> void:
score = 0
text = "Score: %d" % score
func add(points: int) -> void:
score += points
text = "Score: %d" % score
Save and load a dictionary as JSON: pass
The task: Write res://save_game.gd for Godot 4: a script extending RefCounted with two static functions: save_data(data: Dictionary, path: String) -> void writes the dictionary to the file as JSON, and load_data(path: String) -> Dictionary reads it back.
Passed on the first try.
extends RefCounted
class_name SaveGame
static func save_data(data: Dictionary, path: String) -> void:
var file = FileAccess.open(path, FileAccess.WRITE)
if file == null:
push_error("Failed to open file for writing: " + path)
return
var json_string = JSON.stringify(data)
file.store_string(json_string)
file.close()
static func load_data(path: String) -> Dictionary:
var data: Dictionary = {}
if not FileAccess.file_exists(path):
push_error("File does not exist: " + path)
return data
var file = FileAccess.open(path, FileAccess.READ)
if file == null:
push_error("Failed to open file for reading: " + path)
return data
var json_string = file.get_as_text()
file.close()
var json = JSON.new()
var error = json.parse(json_string)
if error != OK:
push_error("Failed to parse JSON at line %d: %s" % [json.get_error_line(), json.get_error_message()])
return data
var parsed_data = json.data
if parsed_data is Dictionary:
data = parsed_data
return data
Health with signals: pass
The task: Write res://health.gd for Godot 4: a script extending Node with health starting at 100, a signal health_changed(old_value: int, new_value: int), a signal died, and a method take_damage(amount: int) that lowers health without going below 0, emits health_changed when health changes, and emits died once, when health reaches 0.
Passed on the first try.
extends Node
signal health_changed(old_value: int, new_value: int)
signal died
@export var health: int = 100
var _is_dead: bool = false
func take_damage(amount: int) -> void:
if amount <= 0:
return
var old_value: int = health
var new_value: int = max(0, health - amount)
if new_value != old_value:
health = new_value
health_changed.emit(old_value, new_value)
if health == 0 and not _is_dead:
_is_dead = true
died.emit()
Fade out with a tween: pass
The task: Write res://fader.gd for Godot 4: a script extending Node2D with a method fade_out(duration: float) that fades the node's modulate alpha from 1 to 0 over duration seconds with a tween, and a signal faded emitted when the fade ends.
Passed on the first try.
extends Node2D
signal faded
func fade_out(duration: float) -> void:
var tween := create_tween()
tween.tween_property(self, "modulate:a", 0.0, duration)
tween.finished.connect(_on_fade_finished)
func _on_fade_finished() -> void:
faded.emit()
Spawn on a timer, then stop: pass
The task: Write res://spawner.gd for Godot 4: a script extending Node that, once in the scene, adds a new child Node2D every 0.25 seconds, using a Timer, and stops after 4 children. It keeps the number spawned in a variable spawned_count.
Passed on the first try.
extends Node
const SPAWN_INTERVAL: float = 0.25
const MAX_SPAWNS: int = 4
var spawned_count: int = 0
var _spawn_timer: Timer
func _ready() -> void:
_spawn_timer = Timer.new()
_spawn_timer.wait_time = SPAWN_INTERVAL
_spawn_timer.one_shot = false
_spawn_timer.autostart = true
_spawn_timer.timeout.connect(_on_spawn_timer_timeout)
add_child(_spawn_timer)
func _on_spawn_timer_timeout() -> void:
if spawned_count >= MAX_SPAWNS:
_spawn_timer.stop()
return
var node := Node2D.new()
add_child(node)
spawned_count += 1
if spawned_count >= MAX_SPAWNS:
_spawn_timer.stop()
A door with states: pass
The task: Write res://door.gd for Godot 4: a script extending Node for a door with the states CLOSED, OPENING and OPEN. It starts CLOSED. A method interact() starts opening a closed door; 0.5 seconds of game time later it is OPEN. A method get_state_name() -> String returns "CLOSED", "OPENING" or "OPEN".
Passed on the first try.
extends Node
enum State {
CLOSED,
OPENING,
OPEN
}
var _state: int = State.CLOSED
func interact() -> void:
if _state != State.CLOSED:
return
_state = State.OPENING
await get_tree().create_timer(0.5).timeout
_state = State.OPEN
func get_state_name() -> String:
match _state:
State.CLOSED:
return "CLOSED"
State.OPENING:
return "OPENING"
State.OPEN:
return "OPEN"
return "CLOSED"
An inventory with counts: pass
The task: Write res://inventory.gd for Godot 4: a script extending RefCounted with add_item(item_name: String, count: int), remove_item(item_name: String, count: int) -> bool that removes only when there are enough and returns whether it did, and get_count(item_name: String) -> int that is 0 for an item never added.
Passed on the first try.
extends RefCounted
var _items: Dictionary = {}
func add_item(item_name: String, count: int) -> void:
if item_name.is_empty() or count <= 0:
return
if _items.has(item_name):
_items[item_name] += count
else:
_items[item_name] = count
func remove_item(item_name: String, count: int) -> bool:
if not _items.has(item_name):
return false
if _items[item_name] < count:
return false
_items[item_name] -= count
if _items[item_name] <= 0:
_items.erase(item_name)
return true
func get_count(item_name: String) -> int:
if not _items.has(item_name):
return 0
return _items[item_name]
Load a level scene from a path: pass
The task: Write res://level_loader.gd for Godot 4: a script extending Node with a method load_level(path: String) -> Node that loads the scene file at path, creates an instance of it, adds it as a child, and returns it.
Passed on the first try.
extends Node
func load_level(path: String) -> Node:
var scene: PackedScene = load(path)
var instance: Node = scene.instantiate()
add_child(instance)
return instance
Paint cells on a TileMapLayer: pass
The task: Write res://grid_painter.gd for Godot 4.7: a script extending TileMapLayer with paint_row(y: int, from_x: int, to_x: int) that sets the cells from (from_x, y) to (to_x, y) inclusive to tile source 0, atlas coordinates (0, 0), and is_painted(cell: Vector2i) -> bool that says whether a cell has a tile.
Passed on the first try.
extends TileMapLayer
func paint_row(y: int, from_x: int, to_x: int) -> void:
for x in range(from_x, to_x + 1):
set_cell(Vector2i(x, y), 0, Vector2i(0, 0))
func is_painted(cell: Vector2i) -> bool:
return get_cell_source_id(cell) != -1
Cast a physics ray: pass
The task: Write res://line_of_sight.gd for Godot 4: a script extending Node2D with first_hit(from: Vector2, to: Vector2) -> Vector2 that casts a physics ray through the 2D world from `from` to `to` and returns the point where it first hits a collider, or `to` when it hits nothing. It is called from physics processing.
Passed on the first try.
extends Node2D
func first_hit(from: Vector2, to: Vector2) -> Vector2:
var space_state := get_world_2d().direct_space_state
var query := PhysicsRayQueryParameters2D.create(from, to)
var result := space_state.intersect_ray(query)
if result.is_empty():
return to
return result.position
A path on a grid around walls: pass
The task: Write res://path_finder.gd for Godot 4: a script extending RefCounted with find_path(size: Vector2i, walls: Array[Vector2i], from: Vector2i, to: Vector2i) -> Array[Vector2i] that returns the shortest path of grid cells from `from` to `to`, both included, on a grid of `size` cells, never entering a wall cell and moving only up, down, left or right.
Passed on the first try.
extends RefCounted
const DIRECTIONS: Array[Vector2i] = [
Vector2i(1, 0),
Vector2i(-1, 0),
Vector2i(0, 1),
Vector2i(0, -1)
]
func find_path(size: Vector2i, walls: Array[Vector2i], from: Vector2i, to: Vector2i) -> Array[Vector2i]:
# Validate that start and goal are inside the grid.
if not _in_bounds(from, size) or not _in_bounds(to, size):
return []
# Build a hash set of walls for O(1) lookup.
var wall_set: Dictionary = {}
for wall in walls:
wall_set[wall] = true
# If start or goal sits on a wall, no path is possible.
if wall_set.has(from) or wall_set.has(to):
return []
# Trivial case: start equals goal.
if from == to:
return [from]
# Breadth-first search. Each cell maps to its predecessor; the start
# uses a sentinel value so reconstruction knows when to stop.
var queue: Array[Vector2i] = [from]
var parent: Dictionary = {from: Vector2i(-1, -1)}
while not queue.is_empty():
var current: Vector2i = queue.pop_front()
if current == to:
return _reconstruct_path(parent, to)
for dir in DIRECTIONS:
var next: Vector2i = current + dir
if not _in_bounds(next, size):
continue
if wall_set.has(next):
continue
if parent.has(next):
continue
parent[next] = current
queue.append(next)
# Goal was never reached.
return []
func _in_bounds(cell: Vector2i, size: Vector2i) -> bool:
return cell.x >= 0 and cell.x < size.x and cell.y >= 0 and cell.y < size.y
func _reconstruct_path(parent: Dictionary, end: Vector2i) -> Array[Vector2i]:
var path: Array[Vector2i] = []
var cell: Vector2i = end
while cell != Vector2i(-1, -1):
path.append(cell)
cell = parent[cell]
path.reverse()
return path
Build and play an animation in code: fail
The task: Write res://blinker.gd for Godot 4: a script extending Node2D that, when it enters the scene, creates an AnimationPlayer child and a looping animation named "blink", 1 second long, that keeps the node visible for the first half of each second and hidden for the second half, and plays it.
Failed on the first try (BENCH FAIL: visible at 0.25, 0.75, 1.25 and 1.75 s was [true, true, true, true], expected [true, false, true, false]) and after one fix (BENCH FAIL: visible at 0.25, 0.75, 1.25 and 1.75 s was [true, true, true, true], expected [true, false, true, false]).
What Godot 4.7.2 printed on the first try:
Godot Engine v4.7.2.stable.mono.official.ed1daf0bf - https://godotengine.org
BENCH FAIL: visible at 0.25, 0.75, 1.25 and 1.75 s was [true, true, true, true], expected [true, false, true, false]Wait for a signal or a timeout: pass after one fix
The task: Write res://waiter.gd for Godot 4: a script extending Node with wait_for(sig: Signal, seconds: float) -> bool, a coroutine that waits until the signal is emitted or `seconds` of game time pass, whichever comes first, and returns true if the signal came first. The signal has no arguments.
Failed on the first try (BENCH FAIL: returned false for a signal 0.2 s away with a 1 s timeout), passed after one fix.
What Godot 4.7.2 printed on the first try:
Godot Engine v4.7.2.stable.mono.official.ed1daf0bf - https://godotengine.org
BENCH FAIL: returned false for a signal 0.2 s away with a 1 s timeoutQuestions
- Is MiniMax M3 good at Godot?
- On these 15 tasks it passed 13 on the first try and 14 after one fix. Each task is small and checked by running it, so the score says whether its Godot 4 code works, not how well it designs a game.
Related
- Beam 1.0 501B A23B on the same tasks
- Claude Haiku 4.5 on the same tasks
- DeepSeek V4 Pro 0813 on the same tasks
- Fugu Max on the same tasks
- GPT-5.4 on the same tasks
- Nemotron 3.5 Super VL on the same tasks
- Qwen3.7 Max on the same tasks
- Claude Haiku 5.5 on the same tasks
- Every verified Godot 4 answer
- Can Mistral Large 4 write working Godot 4 code? 12 of 15 on the first try
- Can Ember-1 write working Godot 4 code? 15 of 15 on the first try
- Can Muse Wonder write working Godot 4 code? 15 of 15 on the first try
- Can GLM 5.2 write working Godot 4 code? 14 of 15 on the first try
- Can Claude Opus 5.5 write working Godot 4 code? 15 of 15 on the first try
Flockbay runs your own agent, Claude or ChatGPT on the official Godot 4.7.2, and gives it Godot's errors the way this test does.
Download Flockbay free