PolicyBench - US Exact Match: leaderboard

United States household tax and benefit calculations without tools. The headline household-impact-weighted exact-match metric allows a one-dollar numeric tolerance and exact binary coverage decisions; households receive equal final weight. Native per-model scored-output counts and symmetric reference exclusions are retained. Forced answer-tool calls disable thinking for affected Claude configurations; JSON-contract rows retain their native reasoning settings. Logged run-state settings remain distinct from registry-derived settings. Continuous score, bounded score, cost and latency columns are separate measurements. Metric: Household-impact-weighted exact match (%). Source: https://policybench.org

Metric: Household-impact-weighted exact match (%). Source: policybench.org. Saturation forecast: Around February 2027. 52 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1GPT-6 Sol95.16#29
2Claude Opus 5.594.07#7
3GPT-6.1 Terra93.14#22
4GPT-5.6 Sol93.08#33
5Claude Fable 5.291.95#1
6GPT-6 Astra91.73#8
7Claude Sonnet 5.591.67#21
8GPT-6 Luna91.5#105
9Claude Fable 5.190.84#17
10Kimi K390.53#59
11GPT-6.1 Sol90.16#12
12Muse Wonder90.05#31
13GPT-5.6 Luna88.7#130
14Gemini 4 Argon88.61#14
15GLM-5.3 Flash88.57#136

No result here: #2 GPT-6.2 Astra, #3 Gemini 4 Carbon, #6 GPT-6.2 Sol.

Interactive version: theaggregate.ai/benchmark?slug=policybench-us-exact-match · How It Works · Data refreshed daily, snapshot 2026-11-17.