IntentGrasp - Gem Set - Toxic Speech: leaderboard

The Toxic Speech domain of the IntentGrasp (2026) Gem Set, hard intent-understanding items that every open-weight model failed in its first run. The score is instance F1 in percent, averaged over four option orders; higher is better.

Metric: Instance F1 (%) on the Toxic Speech (TS) domain of the 470-item Gem Set, mean of four option-order runs, zero-shot, API default settings; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)11.5#99
2Claude Sonnet 4.611#173
3Claude Opus 4.77.67#85
4Gemini 3 Flash (Preview)6.12#152
5GPT-5.4 Nano5.83#651
6GPT-5.45#135
7Gemini 3.1 Flash Lite (Preview)4.83#417
8GPT-5.4 Mini2.5#406
9Claude Haiku 4.52#583

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=intentgrasp-gem-set-toxic-speech · How It Works · Data refreshed daily, snapshot 2026-11-21.