GameASG-Bench: leaderboard
GameASG-Bench (Zhang et al., 2026) asks coding agents to generate complete browser games from a gameplay specification and a pre-declared evaluation interface, then checks source-level compliance and scenario-based behavior with real browser input in headless Chromium. Strict task success rate in percent over 47 tasks, requiring every source-level check and every prerequisite and core behavioral check to pass; each row is a model in a coding harness (Claude Code or Codex CLI) at a stated reasoning effort; higher is better.
Metric: Strict task success rate (%; tasks out of all 47 planned browser-game generation tasks whose delivered index.html passes every L1 source-level check and every applicable L2 prerequisite and core-requirement behavioral check in headless Chromium; one run per task from a clean workspace with full tool access; read from the printed success count as count x 100/47). Source: arxiv.org. Saturation forecast: Around October 2028. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (Max) | 55.32 | #6 (GPT-6 Astra) |
| 2 | Claude Opus 5 (Max) | 51.06 | #20 (Claude Opus 5) |
| 3 | GPT-5.6 Sol (Max) | 44.68 | #29 (GPT-5.6 Sol) |
| 4 | DeepSeek V4 Flash (High) | 40.43 | #254 (DeepSeek V4 Flash) |
| 5 | DeepSeek V4 Flash (Max) | 38.3 | #254 (DeepSeek V4 Flash) |
| 6 | Kimi K3 (Max) | 31.91 | #53 (Kimi K3) |
| 7 | DeepSeek V4 Pro (Max) | 31.91 | #181 (DeepSeek V4 Pro) |
| 8 | GLM-5.2 (Max) | 23.4 | #137 (GLM-5.2) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=gameasg-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.