The AI Divergence Index
Across 25 questions SimulQ has put to its four-model panel, the models agreed on 42% of the points they made. 58% were made by three models or fewer, and 17% by only one. In 25 of 25 results, at least one point was raised by a single model alone.
379 points from 20 Docket rulings and 5 comparison pages, adjudicated up to 23 October 2026. Updated each quarter.
How often the four models disagree
For every question, the same four frontier models (GPT, Gemini, Sonar and Grok) answer independently. SimulQ’s On the Surface sorts every point they make by how many of them made it. The Index counts those points.
| Results | Points | All four | Three | Two | One only | |
|---|---|---|---|---|---|---|
| All published results | 25 | 379 | 42% | 23% | 18% | 17% |
By subject
| Results | Points | All four | Three | Two | One only | |
|---|---|---|---|---|---|---|
| AI tools (comparisons) | 5 | 61 | 43% | 26% | 18% | 13% |
| Legal | 5 | 68 | 44% | 21% | 19% | 16% |
| AI | 2 | 38 | 26% | 29% | 24% | 21% |
| Medical | 2 | 36 | 42% | 22% | 19% | 17% |
| Science education | 2 | 25 | 44% | 24% | 20% | 12% |
| Business | 1 | 21 | 29% | 24% | 24% | 24% |
| Engineering | 1 | 15 | 53% | 20% | 20% | 7% |
| Finance | 1 | 16 | 44% | 25% | 13% | 19% |
| Government | 1 | 22 | 36% | 18% | 14% | 32% |
| Health | 1 | 14 | 57% | 21% | 14% | 7% |
| Information technology | 1 | 15 | 60% | 13% | 13% | 13% |
| Marketing | 1 | 13 | 46% | 15% | 15% | 23% |
| Real estate | 1 | 16 | 50% | 19% | 13% | 19% |
| Sales | 1 | 19 | 32% | 26% | 21% | 21% |
By quarter
| Results | Points | All four | Three | Two | One only | |
|---|---|---|---|---|---|---|
| Q4 2026 | 25 | 379 | 42% | 23% | 18% | 17% |
How it is counted
- Source. Every result published on Docket Central and on the comparison pages, each shown there in full, word for word.
- Unit. One point is one item in a result’s agreement tiers: Agreed by all four, By three, By two or Raised by one.
- What it does not say. It counts how often the models differ, not which model was right; points are counted equally whatever their weight. The sample is small, and most of it is legal and AI-tools questions so far.
- Updates. Figures are recounted whenever a result is published or re-adjudicated, and reported each quarter.