DecideBench: leaderboard
Decision models pick one option key for each of 450 multiple-choice business decisions (agent action review, tool routing, moderation, refunds, release gating, triage), with one worked example per option for every model that can take one (three encoder models run zero-shot); the score is the share of the 450 items each model got right. Unlike general reasoning sets, the items were written so a frontier LLM solves almost all of them, so the board measures how close small dedicated decision models come to that.
Metric: Accuracy on 450 contrastive decisions (%). Source: huggingface.co. Saturation forecast: Rough model projection: around 2026. 27 models tracked. Updated 16 Nov 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Drex Max | 94.4 | |
| 2 | REAX 1 27B | 92.4 | |
| 3 | PPLX-Decider-v1.2 | 92.2 | |
| 4 | vLLM-SR-d3 | 91.6 | |
| 5 | Jev 1.14.0 | 88 | |
| 6 | yev2-35B-a3B | 88 | |
| 7 | yev2-35B-a3B [FP8] | 87.6 | |
| 8 | Jev 1.13.0 | 86.2 | |
| 9 | Clef 27B | 85.3 | |
| 10 | Blink-v0.3-26B-A4B [NVFP4] | 82 | |
| 11 | Drex v2.0 | 80.9 | |
| 12 | yev1-4B | 79.3 | |
| 13 | PPLX-Decider-Flash-v1.2 | 78.2 | |
| 14 | yev0-4B | 78.2 | |
| 15 | Nimble 9B | 75.6 |
Interactive version: theaggregate.ai/benchmark?slug=decidebench · How It Works · Data refreshed daily, snapshot 2026-11-16.