Against the frontier
The same 600 Eurobarometer questions and the same prompt for every model, September 2026. On new Eurostat questions tested in October 2026, Claude Opus 5.5 was ahead of our model on accuracy.
- ModelDirectionAccuracy
- Our model27B, one server 82% +12.2%
- Claude Opus 5 75% +9.2%
- GPT-63 examples 74% +12.3%
- Statistical method 74% +3.2%
- GPT-6 72% +11.5%
- Claude Opus 5.53 examples 71% +12.0%
- Claude Opus 5.5 71% +10.0%
- GPT-5.6 66% -1.3%
- Kimi K3 64% +3.7%
Accuracy is how much closer than the country figure. We tie GPT-6 and Claude Opus 5.5 there, the gaps being inside the margin of error, and lead every model on direction.
What it learned from
- 12.3Mpublished survey results it was trained on
- 27EU countries
- 27Bparameters
- 0individual records: group tables only, all licensed for commercial use
What it does not do
- It predicts how groups differ. You supply the national figure; it splits it by group.
- The panel in this demo runs on a frontier model while ours is being integrated.
- It learned from European survey answers. Your own survey would teach it your customers.
- On surveys from organisations it has never seen, such as HSL’s passenger survey, it is no better than the national figure, so there it answers with that figure.
Sealed test evaluated 24 September 2026: 183,544 group results from 167 questions in the four newest Eurobarometer waves, held out of training. Frontier comparison: 600 cells from 157 of those questions. Sources: European Commission Eurobarometer under CC BY 4.0 and Commission Decision 2011/833/EU. Neither licence implies Commission endorsement.