Vals AI Public Benefits Bench: leaderboard
SNAP food benefit questions built with the Center for Civic Futures and Code for America, spanning 50 states and two territories; 230 of 459 scenarios feed published results. The model talks over several turns with an auditor model playing the client and can search the web, then a judge model grades it on about seven expert-validated criteria per question. The number is the share of criteria met, higher is better.
Metric: Accuracy (%). Source: www.vals.ai. Saturation forecast: Around March 2028. 50 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 | 76.93 | #12 |
| 2 | Claude Fable 5.1 | 74.9 | #9 |
| 3 | Muse Wonder Full | 73.75 | #10 |
| 4 | Claude Opus 5.5 | 70.64 | #3 |
| 5 | Claude Fable 5 | 70.43 | #16 |
| 6 | Muse Wonder | 70.23 | #15 |
| 7 | Gemini 4 Argon | 69.76 | #8 |
| 8 | MiMo-V2.6-Pro | 68.94 | #42 |
| 9 | Hy4 preview | 68.61 | #38 |
| 10 | GLM-5.3 | 68.54 | #39 |
| 11 | Muse Spark 1.2 | 68.47 | #62 |
| 12 | Kimi K3 | 68.2 | #30 |
| 13 | Claude Opus 4.8 | 68.13 | #37 |
| 14 | MiMo-V2.6-Flash | 67.59 | #70 |
| 15 | Claude Sonnet 5.5 | 67.19 | #11 |
No result here: #6 GPT-6 Astra.
Interactive version: theaggregate.ai/benchmark?slug=vals-ai-public-benefits-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.