Precision caliper, engineering drawings, and a glowing measurement path

PE Precision Bench · Frontier update

Grok 4.6 and Qwen3.8 join the benchmark.

Eight models answered the same 320 engineering questions. Grok 4.6 and Qwen3.8 Max are new to the set. DeepSeek V4 Pro still has the highest overall score.

97.1%

Kimi K3 · third overall

95.7%

Grok 4.6 · sixth overall

94.6%

Qwen3.8 Max · eighth overall

PE Precision Bench is 320 sealed multiple-choice questions, 20 in each of 16 engineering disciplines. The model picks an answer and names the governing standard. The overall score is 80% answer accuracy and 20% citation match. Tools were off, temperature was 0, and each question was asked once.

DeepSeek V4 Pro scores 97.8% overall, the highest of the eight. Grok 4.6 is sixth at 95.7%. Qwen3.8 Max is eighth at 94.6%. Answer accuracy stays high across the group. The citation scores do not.

The updated leaderboard

RankModelOverallAnswersStandards
1DeepSeek V4 Pro97.8%99.4%91.7%
2GPT-5.6 Sol97.5%99.7%88.9%
3Kimi K397.1%99.4%88.0%
4Grok 4.596.7%99.4%86.1%
5Claude Sonnet 596.2%98.8%86.1%
6Grok 4.695.7%98.1%86.1%
7Gemini 3.5 Flash95.1%97.8%84.3%
8Qwen3.8 Max94.6%98.1%80.6%

Grok 4.5 scored higher than Grok 4.6

Grok 4.6 scores 95.7% overall. Grok 4.5 scores 96.7%. Both have a standards citation score of 86.1%. The gap is answer accuracy: 98.1% for Grok 4.6 and 99.4% for Grok 4.5.

Kimi K3 was the strongest of the new generation

Kimi K3 is third at 97.1%. Its answer accuracy is 99.4%, and its standards citation score is 88.0%. On the 36 questions that ask for the governing standard, its overall score is 97.6%.

Qwen3.8 Max had more failed responses

Qwen3.8 Max answers 98.1% of the questions correctly. Its standards citation score is 80.6%. 4 of its 320 responses still failed after retries. Those failures count as zeroes. Dropping them would raise the score.

Citation scores vary more than answer scores

Answer accuracy for Grok 4.6, Kimi K3, and Qwen3.8 Max is between 98.1% and 99.4%. Their standards citation scores run from 80.6% to 88.0%. A model can choose the right option and still cite the wrong standard, or add a citation the question does not support.

What the score covers

Each model saw the questions once, with no tools and with temperature set to 0. The score is multiple choice plus a citation check. It does not measure whether a model can do engineering work, and it does not replace the current standard or a licensed engineer's review.

Scores by model and discipline

The benchmark page lists all eight models, the score in each discipline, and the scoring rules.

Open the benchmark

Model references