FORTIS - Skill-Grounded Tool Selection Fail Rate: leaderboard

Share (%) of 1,543 skill-grounded tool-selection queries on which a model selects tools beyond the minimum-privilege set or takes no action. FORTIS (2026) builds 20-skill libraries with hidden privilege levels over email, e-commerce and filesystem domains. Lower is better.

Metric: Fail rate (%), the share of over-privileged or no-action outcomes, on FORTIS Task 2 (skill-grounded tool selection, 1,543 queries: the model gets the assigned skill document and the full tool inventory and must pick the minimum-privilege feasible tool set), temperature 0; lower is better. Source: arxiv.org. Saturation forecast: Not forecast. 10 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3.6 Max45.2#193
2Claude Opus 4.747.4#85
3Gemini 3.1 Pro (Preview)49.2#99
4DeepSeek V4 Flash53#267
5Kimi K2.653.7#192
6Claude Sonnet 4.656.8#173
7GPT-5.4 Mini59.3#406
8Gemini 3 Flash61.2#175
9GPT-5.562.5#54
10GPT-5.466.6#135

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

Interactive version: theaggregate.ai/benchmark?slug=fortis-skill-grounded-tool-selection-fail-rate · How It Works · Data refreshed daily, snapshot 2026-11-21.