Measuring how well AI understands the Danish Building Regulations
Byggeriets AI Benchmark independently benchmarks large language models on their ability to interpret and apply Bygningsreglementet (BR18/BR23).
Twelve frontier models, one Danish rulebook
Every model answers the same set of questions drawn from the Danish Building Regulations.
Top accuracy
Gemini 3.1 Pro Preview currently leads the leaderboard.
90% CI 93.5–99.5%Models evaluated
Frontier models from OpenAI, Anthropic and Google.
6 tested Jun 2026 · 6 tested Oct 2026Benchmark questions
Built from real provisions in the Building Regulations.
Bygningsreglementet (BR18/BR23)Best calibration (ECE)
OpenAI GPT-6 Astra has the lowest calibration error of the twelve.
Expected Calibration ErrorAccuracy spread
Gap between the strongest and weakest model tested.
89.5% – 96.5%Avg. question length
Typical length of a benchmark question, in characters.
Range 200–1800 charactersWhich model gets the Building Regulations right?
A live leaderboard of model accuracy with 90% confidence intervals, plus a calibration explorer.
Look inside the benchmark dataset
Explore the composition of the question set that underlies every result on this site.
Our research, written up
Reports and papers describing our methodology, findings, and their implications.
Latest from Byggeriets AI Benchmark
An independent benchmark, built for the Danish construction industry
Byggeriets AI Benchmark is a research initiative that evaluates how reliably large language models can answer questions grounded in Bygningsreglementet.
More about the projectDeveloped by UCN
In collaboration with TUM
Get new benchmark results in your inbox
A short note whenever we add a model, publish a report, or update the dataset.
By subscribing you agree to receive occasional emails.