JamBench - Theme-Only Generation: leaderboard
JamBench (2026) tests project-level game programming on the Godot engine with real Game Jam projects; outputs are checked deterministically in headless mode, and the board is the percentage of projects that compile and then run for 30 seconds without crashing. Here the model builds a complete 2D game from a jam theme keyword alone, writing every file directly; higher is better.
Metric: L3a runtime pass rate (%): share of generated Godot 4 projects that pass file-integrity and headless compilation checks and then run 30 seconds in headless mode without crashing; Task 1a: only the jam theme keyword is given; 50 real Game Jam themes, three runs each; the model writes a complete 2D Godot 4 project from scratch; direct generation, the model writes every file in one pass; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 78.7 | #94 |
| 2 | GPT-5.4 | 77.3 | #123 |
| 3 | Claude Opus 4.6 | 77.3 | #102 |
| 4 | DeepSeek V4 Pro | 72.7 | #188 |
| 5 | Kimi K2.5 | 72 | #256 |
| 6 | GLM-5 | 70.7 | #260 |
| 7 | Qwen 3.5 397B A17B | 67.3 | #246 |
| 8 | Qwen 3.5 27B | 58.7 | #359 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=jambench-theme-only-generation · How It Works · Data refreshed daily, snapshot 2026-11-13.