WebGameBench: leaderboard
Requirement-to-application coding benchmark (WebGameBench, Zhang et al., 2026): from a frozen structured specification of a browser game, the coding agent builds and serves the application in one attempt, and a Codex CLI evaluator with a GPT-5.4 backend at xhigh reasoning plays the delivered game in a real browser and labels it Excellent, Usable or Unusable. The score is the usable rate, the percent of validly scored artifacts across 111 games labelled Excellent or Usable, with evaluator and infrastructure failures left out of the denominator; higher is better.
Metric: Usable Rate (%; validly scored artifacts labelled Excellent or Usable). Source: arxiv.org. Saturation forecast: Around February 2027. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 76.9 | #75 |
| 2 | Claude Opus 4.6 | 73 | #100 |
| 3 | Gemini 3.1 Pro (Preview) | 63.6 | #91 |
| 4 | GPT-5.5 | 63.6 | #48 |
| 5 | Kimi K2.6 | 61 | #173 |
| 6 | DeepSeek V4 Flash (Non-reasoning) | 59.3 | #254 (DeepSeek V4 Flash) |
| 7 | GLM-5.1 | 52.8 | #175 |
| 8 | GLM-5 | 48.2 | #256 |
| 9 | Claude Sonnet 4.5 | 47.1 | #244 |
| 10 | DeepSeek V4 Pro (Non-reasoning) | 44.1 | #181 (DeepSeek V4 Pro) |
| 11 | Kimi K2.5 | 38.3 | #252 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=webgamebench · How It Works · Data refreshed daily, snapshot 2026-11-12.