The best AI now ties the human score at drawing a floor plan from photos of a flat
Shown about 20 photos of a flat, Gemini 4 Carbon draws its floor plan as well as the people tested. In 2025, people beat every model.
Look through about 20 photos of a flat, then sketch its floor plan: which rooms there are and which doors connect them. That means building a mental map of a space from scattered views, which AI models have found hard. Andon Labs’ Blueprint-Bench scores each plan on how closely its rooms and connections match the real layout.
In the first version, released in 2025, people scored 0.547 and the best model, GPT-5, scored 0.431. Andon Labs said people outperformed every AI system, and later described the original’s model outputs as close to noise. Blueprint-Bench 2 turns it into a longer task: each AI works through 50 apartments in a row and keeps a notepad between them. Scores are rescaled so that random guessing gets 0 and a perfect plan gets 1.
As of 9 November 2026, Gemini 4 Carbon scores 0.586, exactly the human baseline. Next come Gemini 4 Argon (0.544), Claude Opus 5.5 (0.512) and GPT-6 Astra (0.497). The spread is wide: Claude Opus 5 scores 0.304, and five models, including Claude Haiku 4.5 and Gemini 3 Flash, still score no better than random.
What it means
The best models can now piece together the layout of a home from scattered photos about as well as the people tested, which none could a year earlier. Smaller and cheaper models still fail at it, so one model’s result says little about another. This is one lab’s test, with the human score taken from only 12 flats.
Caveats
- The human baseline was measured on only 12 apartments, while models are scored on 50.
- Version 1 and version 2 use different scoring scales, so their numbers can’t be compared directly.
- Andon Labs designs, runs and scores the benchmark. Results have not been independently replicated.
- The test checks how rooms connect, not exact sizes or distances.