69.4%
#1
MortgageTax accuracy of Gemini 3.1 Pro Preview, the highest of the top 15; GPT 4.1 was 15th at 65.9%

What the chart shows

MortgageTax tests whether models can read real mortgage tax certificates, as text and document images, and pull out fields such as year, parcel number and county, or compute the annualised amount due. Gemini 3.1 Pro Preview led at 69.4%, followed by Gemini 3 Pro (69.1%) and Gemini 2.5 Pro (68.9%).

Gemini models take the top four places. The top 15 span just 3.5 percentage points, down to GPT 4.1 at 65.9%, and none reaches 70%.

Key numbers

MeasureValue
Gemini 3.1 Pro Preview69.40%
Gemini 3 Pro (11/25)69.08%
Gemini 2.5 Pro68.92%
Gemini 3 Flash (12/25)68.72%
GPT 4.1, 15th65.94%
Gap, 1st to 15th3.46 percentage points

How to read it

  • The dataset holds 1,258 documents; semantic questions ask for fields, numerical ones for the annualised amount due.
  • The bars start at zero, so the narrow spread shows as near-equal bars.
  • Scores are from Vals.ai’s MortgageTax leaderboard; the AI Index shows the top 15 models.
Source: Vals.ai, 2026, via the AI Index Report 2026 (Stanford HAI), figure 2.5.9: MortgageTax: accuracy. Chart and text by AI Adoption.
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart