GBQA: leaderboard
GBQA (Game Benchmark for Quality Assurance, 2026) asks an LLM-driven agent to find the 124 human-verified bugs implanted in 30 generated games within 500 interaction steps per game, with access to design documents and source code (Quality Assurance Mode). Recall (%) of the ground-truth bugs, matched by a critic agent; higher is better.
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 May 2027. 15 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Claude Opus 4.6 | Thinking | 48.39 | #113 |
| 2 | Qwen 3.5 397B A17B | Thinking | 41.13 | #264 |
| 3 | DeepSeek R1 | 37.9 | #482 | |
| 4 | Claude Sonnet 4.5 | Thinking | 37.1 | #261 |
| 5 | Qwen 3 235B A22B | Thinking | 35.48 | #628 |
| 6 | O3 | 34.68 | #237 | |
| 7 | Qwen 3 32B | Thinking | 33.87 | #972 |
| 8 | Gemini-3-Pro [GBQA checkpoint unspecified] | 33.06 | ||
| 9 | Kimi K2.5 | Thinking | 28.23 | #266 |
| 10 | Qwen 3 8B | Thinking | 24.19 | #1873 |
| 11 | GPT-5.2 | Non-reasoning | 22.58 | #185 |
| 12 | DeepSeek V3.2 | Non-reasoning | 20.16 | #419 |
| 13 | Gemini 3 Flash | Non-reasoning | 19.35 | #175 |
| 14 | Llama 3.1 70B Instruct | 14.52 | #1571 | |
| 15 | Llama 3.1 8B Instruct | 8.87 | #4600 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=gbqa · How It Works · Data refreshed daily, snapshot 2026-11-21.