RESULTS

GameCraft-Bench: measured by validation.

GameCraft-Bench evaluates AI agents on 140 real Godot game-development tasks. Agents are ranked by raw score; each score reflects the final effective attempt per task, evaluated by hidden Godot validation scripts and a multimodal judge.

LEADERBOARD

GameCraft-Bench Overall Score

SEELE02-pro row uses the final effective validation results from the Seele report. External rows follow the GameCraft-Bench leaderboard reference.

RankModelHarnessFeedbackCasesScore
1SEELE02-proSeeleAgentBaseline140
43.8%
2Opus-4.7 highClaude CodeBaseline
41.5%
3GPT-5.5 highCodexBaseline
39.5%
4Kimi-K2.6Kimi CodeBaseline
30.7%
5MiMo-V2.5-ProClaude CodeBaseline
24.1%
6GLM-5.1Code BuddyBaseline
18.3%
7MiniMax-M2.7Code BuddyBaseline
11.0%
8DeepSeek-V4-ProCodexBaseline
2.2%

KEY FINDINGS

What the numbers say in one read.

The headline takeaways from the leaderboard above — every figure traces back to the official GameCraft-Bench evaluation.

43.8%raw score on 140 cases

SEELE02-pro tops GameCraft-Bench with a 43.8% overall score — 2.3 points ahead of the nearest competitor (Claude Code / Opus-4.7, 41.5%).

140 / 140fully scored

All 140 cases were completed and scored in the final effective run — full coverage of the benchmark with no gaps.

87.2%best single case

The highest-scoring case hit 87.2%, demonstrating the model can handle complex Godot tasks with high fidelity.

100%verifier-confirmed

Every score comes from the official verifier pipeline and hidden rubric — model self-reporting is never counted as evidence of success.

WHAT IS GAMECRAFT-BENCH?

A real game-development benchmark, not a code-only test.

GameCraft-Bench is an open-source benchmark of 140 real Godot game-development tasks across 15 game families — from platformers and strategy to roguelikes and visual novels. Each task asks an agent to turn a natural-language game specification into a complete, playable Godot project. Unlike pass/fail coding benchmarks, every case is graded on a percentage scale by the official verifier pipeline and a multimodal judge, so the overall score measures how completely an agent ships a working game — not just whether code compiles.

140real Godot tasks
15game families
2evaluation layers per case
1final effective attempt per task

Verifier pipeline

Each submission is exported as a complete, runnable Godot project. An automated verifier replays recorded demo traces, and a multimodal judge scores the result against a hidden rubric.

Scoring protocol

Overall score is the raw sum of per-case scores divided by total cases. Each case is scored by a hidden rubric; model self-reporting is never counted as evidence of success.

EXAMPLE TASKS

What the benchmark actually asks agents to do.

These examples are compressed from real GameCraft-Bench task descriptions, preserving the actual task substance while making them readable on the page.

Gameplay logictask_001

Inventory system with save/load

Implement a persistent inventory system that saves item state to disk and restores it across sessions, using Godot’s ResourceSaver and ResourceLoader.

UI scene constructiontask_002

HUD with dynamic minimap

Build a heads-up display with a real-time minimap rendered via SubViewport, player health bar, and ammo counter using Control nodes and custom drawing.

Animation / runtime statetask_003

State-machine-driven character

Wire an AnimationTree with a BlendSpace2D for directional movement, transition guards for idle/run/jump states, and call-method tracks for footstep SFX.

Physics / interactiontask_004

Rigid-body stacking puzzle

Create a physics puzzle where the player drags and stacks RigidBody3D crates to reach a trigger zone, with collision filtering and stability thresholds.

Shader / material setuptask_005

Dissolve damage effect

Write a spatial shader with a noise-driven dissolve threshold, edge glow, and color ramp that triggers on enemy health change via uniform updates.

TileSet / level designtask_006

Procedural dungeon generator

Use TileMapLayer with custom tile data to generate a bounded dungeon from a walker algorithm, placing walls, floors, and chests with collision metadata.

SOURCES

Evidence and references

Your studio is one prompt away.

SeeleAgent builds 2D and 3D games — mobile and multiplayer — with export to Unity 6, Three.js, and WebGL, and distribution to Steam, App Store, and Google Play. Open the workspace and ship your first game, video, or scene tonight.