AsgardBench: leaderboard
Metric: Task success rate (%) with image observations (the previous and current views), on AsgardBench's 108 AI2-THOR household tasks (12 task types with varied object states and placements) in which the agent issues high-level actions with navigation and grasping abstracted away and must adapt its plan to what it observes, receiving only simple success or failure feedback; temperature 0, median over independent runs; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.5 (High) | 76.5 | #79 (Claude Opus 4.5) |
| 2 | Gemini 3 Pro (Preview) (High) | 72.5 | #64 (Gemini 3 Pro (Preview)) |
| 3 | GPT-5.2 (High) | 71.3 | #105 (GPT-5.2) |
| 4 | Kimi K2.5 | 68.8 | #139 |
| 5 | Qwen 3 VL 235B A22B (Thinking) | 32.4 | #228 (Qwen 3 VL 235B A22B) |
| 6 | GLM-4.6V | 30.6 | #309 |
| 7 | GPT-4o | 23.4 | #333 |
| 8 | Mistral Large 3 | 7.4 | #388 |
| 9 | Llama 4 Maverick Instruct FP8 | 5.8 | #403 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=asgardbench · How It Works · Data refreshed daily, snapshot 2026-10-11.