GBQA: leaderboard
Metric: Recall (%) in Quality Assurance Mode (the agent also reads the game's design documents and source code) on the 124 human-verified bugs implanted in GBQA's 30 games, found by the model driving the authors' ReAct QA agent (with in-session and cross-session memory) for at most 500 interaction steps per game; a GPT-5.2 critic agent matches each report to the ground-truth bugs; higher is better. Source: arxiv.org. Saturation forecast: Around December 2027. 22 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.6 (Thinking) | 48.39 |
| 2 | Qwen 3.5 397B A17B (Thinking) | 41.13 |
| 3 | Claude Opus 4.6 | 37.9 |
| 4 | DeepSeek R1 | 37.9 |
| 5 | Claude Sonnet 4.5 (Thinking) | 37.1 |
| 6 | Qwen 3 235B A22B (Thinking) | 35.48 |
| 7 | O3 | 34.68 |
| 8 | Qwen 3 32B (Thinking) | 33.87 |
| 9 | Claude Sonnet 4.5 | 32.26 |
| 10 | Kimi K2.5 (Thinking) | 28.23 |
| 11 | Qwen 3 8B (Thinking) | 24.19 |
| 12 | Qwen 3.5 397B A17B (Non-reasoning) | 24.19 |
| 13 | GPT-5.2 (Non-reasoning) | 22.58 |
| 14 | Kimi K2.5 (Non-reasoning) | 20.97 |
| 15 | DeepSeek V3.2 (Non-reasoning) | 20.16 |
Interactive version: theaggregate.ai/benchmark?slug=gbqa · How It Works · Data refreshed daily, snapshot 2026-10-07.