Togal Released Construction AI Benchmark Test
The new Construction Last Exam benchmark shows that current AI models struggle to interpret complex PDF engineering plans.
Updated on Oct. 13, 2026 in Artificial Intelligence
Live Poll
Do you trust current general-purpose AI models to accurately handle complex technical professional tasks?
Togal has unveiled the Construction Last Exam, a benchmark designed to test artificial intelligence models on 33 technical construction takeoff tasks. Initial results indicate that frontier AI models from major providers achieved scores as low as 2% when interpreting raw PDF drawings.
Why it matters
The testing highlights a significant performance gap for general-purpose AI models tasked with precise technical drawing interpretation. This benchmark aims to quantify the difficulty these systems face when identifying specific elements and producing accurate measurements without third-party assistance.
The benchmark includes 33 specific tasks requiring models to interpret raw PDF drawings using built-in vision and coding capabilities. Performance is measured against ground-truth data using intersection over union, F1, and mean average precision metrics.
The players
Togal
A Miami-based company that develops machine-vision models trained on millions of construction plans.
The details
Models from OpenAI, Anthropic, Google, xAI, and various Chinese providers were tested on their ability to identify elements and extract measurements from plans. The evaluation exclusively utilized the native capabilities of the models without integrating external software tools.
Timeline
The findings regarding the construction benchmark were published on October 13, 2026.
The Tech Race
The Construction Last Exam benchmark represents a shift in evaluating how general-purpose AI models handle specialized technical industry domains. This standardized testing positions the industry to move beyond general chatbot performance toward proving competence in complex, high-stakes professional engineering tasks.
Construction professionals and software developers should note that current top-tier AI models lack the accuracy required for autonomous project takeoff tasks. Users relying on these tools for technical measurement must independently verify outputs to avoid significant engineering errors.
The takeaway
The performance gap seen in this test suggests that general AI remains unsuited for replacing specialized construction takeoff software. Developers should focus on domain-specific training to bridge the gap between general reasoning and precise technical drawing interpretation.
Further reading
For more on the current state of industry-specific models, see our Artificial Intelligence section.
Source note: This article includes information reported by Construction Briefing.
Live Poll
Do you trust current general-purpose AI models to accurately handle complex technical professional tasks?
