GenXR-Bench: leaderboard
Coding agents (mini-swe-agent) build or fix extended-reality apps on the open-source XR Blocks WebXR framework, and held-out tests drive a simulated headset to check the result; the score is the share of 112 tasks passed in one attempt. Unlike SWE-bench-style repository tasks, passing depends on runtime behaviour under simulated hand, head and sensor input.
Metric: Pass@1 over all 112 XR development tasks: share passed in one attempt with every held-out test passing (%). Source: arxiv.org. Saturation forecast: Around May 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 (Medium) | 50.9 | #22 (Claude Opus 5) |
| 2 | GPT-5.6 Sol (xHigh) | 44.6 | #30 (GPT-5.6 Sol) |
| 3 | Gemini 3.8 Flash (High) | 42.9 | #36 (Gemini 3.8 Flash) |
| 4 | Gemini 3.7 Flash (High) | 41.1 | #32 (Gemini 3.7 Flash) |
| 5 | Gemini 3.6 Flash (High) | 31.2 | #76 (Gemini 3.6 Flash) |
| 6 | GPT-5.6 Luna (xHigh) | 30.4 | #121 (GPT-5.6 Luna) |
| 7 | Gemini 3.5 Flash (High) | 29.5 | #101 (Gemini 3.5 Flash) |
| 8 | Gemini 3.1 Pro (Preview) (High) | 23.2 | #94 (Gemini 3.1 Pro (Preview)) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=genxr-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.