GTA-Workflow - Perception: leaderboard
Metric: Root success rate (%) on the tasks that require perception tools on GTA-Workflow, the open-ended long-horizon workflow tier of GTA-2 (132 real-user tasks with deployed tools and multimodal inputs), run end to end in the default Lagent ReAct framework; deliverables judged by GPT-5.2 against recursive checkpoint rubrics; higher is better. Source: arxiv.org. Saturation forecast: Around 2028. 13 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Qwen 3 235B A22B | 15.79 |
| 2 | Llama 4 Scout | 15.79 |
| 3 | Gemini 2.5 Pro | 13.16 |
| 4 | GPT-5 | 13.16 |
| 5 | Claude Sonnet 4.5 | 10.53 |
| 6 | DeepSeek V3.2 | 10.53 |
| 7 | Kimi K2 | 10.53 |
| 8 | Grok 4 | 7.89 |
| 9 | Llama 3.1 70B Instruct | 2.63 |
| 10 | Qwen 3 30B A3B | 2.63 |
| 11 | Llama 3.1 8B Instruct | 0 |
| 12 | Qwen 3 8B | 0 |
| 13 | Llama 3.2 3B Instruct | 0 |
Interactive version: theaggregate.ai/benchmark?slug=gta-workflow-perception · How It Works · Data refreshed daily, snapshot 2026-10-07.