Immigration Bench
Professional Work · 2026-10-28
Vals AI and the Talent Mobility Fund's Immigration Bench v1 asks 262 private expert-written U.S. immigration questions, each from one of seven personas (students, scholars, university advisors, employers, entrepreneurs, self-petitioners and people with fragile or temporary status). The model talks with a GPT-6 Luna simulated person for up to five exchanges and a Claude Sonnet 5 grader marks each expert rubric item met if it appears anywhere in the conversation. The score is average answer accuracy: the share of a question's rubric items met, averaged over questions. Atlas imports the leaderboard run without web search; a conversation over its 30-minute limit scores zero.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5.1 | 75.4 |
| Claude Opus 5.5 | 72.1 |
| Muse Wonder Full | 71.2 |
| Muse Wonder | 69.6 |
| Claude Sonnet 5.5 | 69.0 |
| Kimi K3 | 68.3 |
| Gemini 4 Argon | 68.1 |
| Grok 4.7 | 68.0 |
| GPT-6.1 Sol | 67.8 |
| GPT-6 Astra | 67.7 |
| MiMo-V2.6-Pro | 67.5 |
| GLM-5.3 | 66.7 |
| Muse Spark 1.2 | 66.5 |
| Grok 4.6 | 66.0 |
| DeepSeek V4.1 Flash | 65.6 |
| Muse Spark 1.3 | 65.0 |
| Mistral Large 4 Preview | 63.7 |
| Gemini 3.8 Flash | 63.4 |
| GPT-6 Luna | 62.5 |
| Claude Haiku 5.5 | 61.8 |
| DeepSeek V4 Pro 0813 | 61.7 |
| MiniMax M3 | 57.5 |
| Gemini 3.5 Flash-Lite | 55.7 |