GBENCH One-shot Coding GScore
Games · 2026-04-06
GBENCH One-shot Coding evaluates models by giving them complete descriptions of novel game environments and allowing one code submission with no retries. GScore combines percentile performance across games, successful execution, and a small response-time penalty on a 0-100 scale.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 79.2 |
| GPT-6.1 Sol | 79.0 |
| GPT-6.1 Terra | 77.6 |
| Claude Opus 5.5 | 73.9 |
| GPT-6 Sol | 70.1 |
| Claude Fable 5.1 | 69.3 |
| Claude Opus 5 | 68.1 |
| Gemini 4 Argon | 67.9 |
| GPT-5.6 Sol | 67.7 |
| Claude Sonnet 5.5 | 67.4 |
| Claude Fable 5 | 65.6 |
| GPT-5.6 Terra | 63.3 |
| GPT-5.5 | 62.2 |
| Gemini 3.8 Flash | 58.6 |
| Kimi K3.1 Pro | 58.2 |
| Muse Wonder | 58.1 |
| GPT-5.4 | 56.5 |
| Gemini 3.6 Flash | 56.3 |
| Grok 4.7 | 56.3 |
| Claude Haiku 5.5 | 55.6 |
| Grok 4.6 | 55.4 |
| Gemini 3.1 Pro Preview | 54.9 |
| Gemini 3.5 Flash | 54.9 |
| Muse Spark 1.1 | 54.3 |
| GPT-6 Luna | 54.1 |
| Kimi K3 | 54.0 |
| Opus 4.8 | 54.0 |
| DeepSeek V4.1 Flash 1017 | 53.8 |
| Ember-1 | 53.7 |
| Mercury 3 | 53.5 |
| MiMo-V2.6-Pro | 53.4 |
| Muse Spark 1.3 | 52.7 |
| Gemini 3.7 Flash | 52.6 |
| GPT-5.3-Codex | 50.2 |
| GPT-5.6 Luna | 50.2 |
| Gemini 3 Pro Preview | 50.0 |
| MiMo-V2.6-Flash | 49.6 |
| GLM-5.3 Flash | 49.3 |
| Opus 4.6 | 48.9 |
| DeepSeek V4 Pro 0813 | 48.7 |