2.9 pts
gap between the best (Claude Sonnet 4.6, 77.1%) and 15th-placed (Kimi K2.5, 74.2%) models on TaxEval v2

What the chart shows

TaxEval v2 contains over 1,500 expert-verified tax questions covering numerical reasoning, problem solving and compliance rules, and scores both the answer and the reasoning behind it. Claude Sonnet 4.6 led the top 15 at 77.1%, ahead of Claude Opus 4.6 (Thinking) at 76% and Grok 3 at 75.9%.

The whole top 15 lies within 2.9 percentage points, down to Kimi K2.5 at 74.2%. Older models such as GPT 4o (74.5%) and o1 (74.3%) remain within the band.

Key numbers

MeasureValue
Claude Sonnet 4.677.11%
Claude Opus 4.6 (Thinking)75.96%
Grok 375.88%
GPT 4o (2024-11-20)74.53%
Kimi K2.5, 15th74.20%
Gap, 1st to 15th2.91 percentage points

How to read it

  • The report text names Claude 3.7 Sonnet (Thinking) at 74% as the lowest of the 15; it is not in the published data, where the lowest is Kimi K2.5 at 74.2%. We use the data.
  • The bars start at zero, so they look almost equal: that is the finding.
  • Scores are from Vals.ai’s TaxEval v2 leaderboard; the AI Index shows the top 15 models.
Source: Vals.ai, 2026, via the AI Index Report 2026 (Stanford HAI), figure 2.5.8: TaxEval v2: accuracy. Chart and text by AI Adoption.
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart