Overall score
Who performed best across the whole test?
80% answer accuracy + 20% standards score.
Independent model comparison
How well can today's AI answer professional engineering questions?
We gave five leading AI models the same 320-question engineering test. We checked both whether they chose the right answer and whether they used engineering standards correctly.
320 questions16 disciplines5 AI models
The short version
The models were excellent at choosing answers. Correctly citing standards was harder.
97.8%–99.7%
Range of correct-answer rates across all five models.
84.3%–91.7%
Range of standards scores—the clearest difference between models.
Very close race
Every overall score fell between 95% and 98%, so discipline details matter.
How to read the scores
Three numbers, three simple questions
Answers correct
Did the model choose the right engineering answer?
Higher is better.
Standards score
Did it name the right code or standard without adding unsupported ones?
Higher is better.
Overall results
The five models, ranked
The bar scale starts at 90% so the small differences are easier to see.
1
DeepSeek V4 Pro
WinnerDeepSeek
Overall97.8%
Answers correct99.4%
Standards score91.7%
Unsupported citations23.3% · Some extra citations
2
GPT-5.6 Sol
OpenAI
Overall97.5%
Answers correct99.7%
Standards score88.9%
Unsupported citations15.8% · Most restrained citations
3
Grok 4.5
xAI
Overall96.7%
Answers correct99.4%
Standards score86.1%
Unsupported citations29.5% · Some extra citations
4
Claude Sonnet 5
Anthropic
Overall96.2%
Answers correct98.8%
Standards score86.1%
Unsupported citations50.8% · Frequent extra citations
5
Gemini 3.5 Flash
Overall95.1%
Answers correct97.8%
Standards score84.3%
Unsupported citations46.9% · Frequent extra citations
Important: “Unsupported citations” means the model named standards that were not required or supported by the expected answer. Lower is better. A correct final answer can still include an unreliable reference.
Explore by discipline
Results for your kind of engineering
Choose a discipline to compare the same five models on its 20 questions.
Best result in Agricultural and Biological
GPT-5.6 Sol
100.0%
1GPT-5.6 Sol
Overall100.0%
Answers100.0%
Standards100.0%
2Claude Sonnet 5
Overall100.0%
Answers100.0%
Standards100.0%
3Gemini 3.5 Flash
Overall96.0%
Answers95.0%
Standards100.0%
4Grok 4.5
Overall96.0%
Answers95.0%
Standards100.0%
5DeepSeek V4 Pro
Overall96.0%
Answers95.0%
Standards100.0%
Discipline-level standards scores may be based on only a few standards-specific questions, so use them as directional signals rather than definitive rankings.
See a real example
Try one sample question
These four public examples are not part of the scored test. The other 320 questions stay private so future models cannot memorize them.
Civil•Medium
A full reservoir discharges through a 0.50 m² outlet at a point 12 m below the free surface. Neglecting losses, what is the approximate outlet velocity?
How we ran the test
Same questions. Same rules. No outside help.
Same test
Every model answered the same 320 questions.
Closed book
Web search and external tools were turned off.
One attempt
Each model got one zero-temperature run per question.
Private holdout
The scored questions and model answers remain sealed.
Technical details and reproducibility record
Scoring80% correct answers plus 20% standards citation score.
Coverage284 analysis questions and 36 standards questions.
Run cost$2.75 provider-reported successful-response usage; $8.70 conservative estimate.
PublishedJuly 15, 2026 · benchmark v1.0.0
Sealed dataset SHA-2568499cf66c3bf66541641f18d667d8c9d0726770849ff5787db28f2de25ac2c3a
What should you take away?
AI models can be remarkably accurate on structured engineering questions, but a correct answer does not guarantee a trustworthy standards citation. Always verify governing codes and project decisions with current source documents and accountable professional review.
Read the benchmark story