Werdykt v1 - Category Abstain: leaderboard
Werdykt v1 Category Abstain measures text-only decisions in Polish and English. The full test has 5,000 hidden questions in twenty balanced language/category groups of 250; category panels combine 500 questions and language panels have 2,500. Macro accuracy gives groups equal weight. Request errors and unparsed answers count incorrect. Native non-control systems retain their interface, engine, reasoning, legacy and source-independence flags; decision endpoints remain distinct from chat.
Metric: Accuracy (%). Source: basal.si5.pl. Saturation forecast: Rough model projection: around 2026. 121 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.8 Flash (Werdykt Run: native=Gemini%203.8%20Flash; cfg=d6e25ef16f84a68f) | 100 | |
| 2 | GPT-6.1 Sol (Werdykt Run: native=GPT-6.1%20Sol; cfg=9bc38d71de848a6b) | 100 | |
| 3 | GPT-6 Astra (Werdykt Run: native=GPT-6%20Astra; cfg=4a80c27c0021a614) | 100 | |
| 4 | Claude Opus 5.5 (Werdykt Run: native=Claude%20Opus%205.5; cfg=6d04e9478761c26f) | 99.8 | |
| 5 | GPT-6 Luna (Werdykt Run: native=GPT-6%20Luna; cfg=b572761faa59abee) | 99.6 | |
| 6 | Gemini 3.1 Pro (Preview) (Werdykt Run: native=Gemini%203.1%20Pro%20%28preview%29; cfg=e8ffbba3cfe39393) | 99 | |
| 7 | Liquid d1-ultra (Werdykt Run: native=Liquid%20d1-ultra; cfg=4faf298d55c875bb) | 93 | |
| 8 | DREX Max (OpenRouter Decisions API) (Werdykt Run: native=DREX%20Max%20%28OpenRouter%20Decisions%20API%29; cfg=62a8151386517b05) | 92.4 | |
| 9 | mammoth-1 (Werdykt Run: native=mammoth-1; cfg=7ac9b4868f7083b8) | 90 | |
| 10 | pplx-decider-v1-27B (Werdykt Run: native=pplx-decider-v1-27b; cfg=dd1d846fc0af2505) | 87.2 |
Interactive version: theaggregate.ai/benchmark?slug=werdykt-v1-category-abstain · How It Works · Data refreshed daily, snapshot 2026-11-05.