Texas moved its electrician exams to the 2026 National Electrical Code on September 1, 2026. Three weeks later we ran our chatbot test on the Texas Journeyman Electrician exam: the same nine-prompt conversation, six models (GPT-6 Astra, Claude Fable 5, Gemini 3 Pro, DeepSeek V4, Grok 4, Kimi K3), the same questions a candidate would ask. What do I need? Who gives the exam? How many questions? Which code book? Can I tab it? Then a 20-question mock with at least five calculations.
We graded every fact against the PSI Candidate Information Bulletin dated September 15, 2026 and the TDLR pages, and every mock answer key against the actual text of the 2026 NEC.
Here's the headline: five of six knew the exam had switched to the 2026 code. Not one of them had read it.
What the exam actually is
Since March 2025 the Texas journeyman exam is two separately-passed portions. NEC Knowledge: 59 items, 3 of them unscored, 130 minutes. Calculations: 26 items, 2 unscored, 110 minutes. 70 percent on each. $78 covers both portions; a retake is $78. Open book, and the only book is a soft-bound 2026 NEC. Publisher-made tabs only. You may highlight, underline and write notes in it before the exam. Nothing goes in during. No Handbook, no Ugly's, no paper. TDLR lets you sit the exam at 7,000 hours of supervised on-the-job training and issues the license at 8,000. The application fee is $30.
One more from TDLR's own statistics page: in fiscal 2025, 24.46 percent passed the NEC portion and 20.56 percent passed Calculations. Hold those, because they're the numbers the bots got wrong, and the reason it matters.
The scorecard
| Model | Facts about the exam and license (8 asked) | Mock: right / wrong / flawed | Worst error |
|---|---|---|---|
| DeepSeek V4 | 6.5 | 13 / 1 / 6 | Keyed the dwelling lighting load at 3 VA per square foot. The 2026 code says 2. |
| Kimi K3 | 5.5 | 16 / 0 / 4 | Said the current bulletin "still references the 2023 NEC." It doesn't. |
| GPT-6 Astra | 5 | 16 / 2 / 2 | "PSI supplies the code book. You generally cannot bring your own." |
| Claude Fable 5 | 4.5 | 11 / 2 / 7 | Tabs are "generally permitted, add more or reorganize them." |
| Gemini 3 Pro | 3 | 18 / 0 / 2 | 80 questions, 240 minutes, 56 to pass. That exam ended in March 2025. |
| Grok 4 | 2 | 3 / 16 / 1 | "As of September 2026, the exam is based on the NEC 2023 edition." |
"Flawed" means the answer was right but the explanation cited a section number that no longer exists in the 2026 book, or taught something the code contradicts.
They knew the edition changed. They didn't know what changed.
Ask the six "which edition is the exam on right now?" and five say 2026 NEC, effective September 1. Good. Then read their mocks.
Five of the six cited 2023 section numbers in their mocks; the sixth cited nothing at all. Table 220.55 for range demand. Table 300.5 for burial depth. 200.6 for neutral identification. 406.4(D) for replacement receptacles. In the 2026 code, load calculations moved from Article 220 to Article 120, burial depth is Table 300.7(A), neutral identification is 200.7. Across 120 mock questions, zero citations to Article 120. Two models told the candidate to put a tab on "Article 220." There is no Article 220 to tab.
That's a lookup problem, and on an open-book exam with 130 minutes for 59 questions it's the whole game. But one of them was worse than a bad citation. DeepSeek, which had the best fact answers of the six, keyed a dwelling general-lighting calculation at 3 volt-amperes per square foot: 2,000 square feet, 6,000 VA. That was the 2023 number. NEC 2026 section 120.41 sets it at 2 VA per square foot. The right answer, 4,000 VA, wasn't among the four choices. A candidate who learns that question walks into the Calculations portion carrying the old number into every dwelling-load problem on the test.
The book rules, one paragraph long, six different answers
The bulletin's reference-material section is short. Soft-bound NEC only. Publisher tabs only. Highlight, underline and write notes before the exam all you want. Nothing added during. Here's what the bots said when asked directly:
- GPT: PSI provides the code book at the center; you can't bring yours. Also, no handwritten notes. Both backwards.
- Gemini: Pre-printed tabs from an exam-prep company are fine (they aren't), and if the proctor finds handwritten notes you can't use the book (you can).
- Fable: Tabs generally permitted, feel free to add more. Publisher tabs only; a DIY-tabbed book gets pulled at the door.
- Grok: No notes, tabs or markings. Then, two lines later, you can tab and highlight. Bring your 2023 book.
- Kimi: First answer said bring it unmarked, later answers got it right.
- DeepSeek: Got the entire policy right.
One in six. The bulletin's consequence for showing up with a non-compliant book is testing without it or forfeiting the $78.
The breakdown nobody looked up
Prompt four asked for the section breakdown. The bulletin prints it: ten subject areas with item counts for each portion. Kimi reproduced it exactly, all twenty numbers. GPT gave a generic topic list. The other four invented one. Gemini produced nine areas totaling 80 questions "according to the official PSI CIB," with 22 wiring-methods questions (real number: 10) and calculations "integrated throughout" (they're a separate portion). DeepSeek gave percentages "exam-prep providers consistently report" for categories that don't exist on the outline. Fable said it didn't have the data, said it would pull it up, then produced two tables with weights to two decimal places. Grok listed topics that aren't on the outline at all.
Three of the six also volunteered a pass rate: Kimi said 24 percent on Knowledge and 21 percent on Calculations for fiscal 2025, Fable said 20 to 25, Grok said 19 to 22. Our first grading pass flagged all three as invented, because our answer key didn't have a pass rate in it. Then we found TDLR's exam statistics page. Fiscal 2025: Journeyman NEC 24.46 percent, Journeyman Calculations 20.56 percent. Kimi was exact. We were the ones who hadn't looked it up, which is the failure mode this whole series is about, so it stays in the post.
That number matters more than anything else on this page. Three in four people who sit the Texas journeyman exam fail it.
The mocks: accuracy and difficulty had nothing to do with each other
We scored each mock for guessability the same way we score our own pools every night: how many of the 20 can a person with no electrical training answer from the choices alone.
| Model | Answerable with no electrical knowledge | Need the code book open |
|---|---|---|
| Grok 4 | 3 of 20 | (16 keys wrong) |
| Gemini 3 Pro | 4 of 20 | 16 of 20 |
| DeepSeek V4 | 4 of 20 | 13 of 20 |
| Kimi K3 | 7 of 20 | 9 of 20 |
| Claude Fable 5 | 13 of 20 | 4 of 20 |
| GPT-6 Astra | 14 of 20 | 2 of 20 |
Gemini wrote the best mock in the group, 18 of 20 verified against the 2026 text and 16 that force you into the book, and paired it with the worst fact answers. GPT's mock was 16 of 20 accurate and 70 percent guessable: Ohm's-law arithmetic, "what device protects a circuit," distractors like "a second neutral." Two items needed the code. Fable's six calculation questions printed every table value in the stem, so the candidate never opens a table, which is the skill the 110-minute Calculations portion is timed against. Grok's five calculation keys were all wrong, and none of them followed from the numbers in its own stems.
For reference, our Texas Knowledge pool scores about 22 percent guessable on that same nightly check and the Calculations pool about 10 percent.
The one we ship
Fable 5 is the same model family we generate our own questions with. It finished fourth of six on facts: right on the two-portion format, the 2026 date and the 7,000-hour rule, wrong on tabs when asked directly, and it fabricated the section weights. Its answer key reversed itself mid-explanation three times, with the "wait" left in the text. We said we'd grade the family we use. Here is the grade.
It's also why the pipeline exists. A model answering from memory cited Article 220 for a 2026 exam. Our generator doesn't answer from memory; it writes from the extracted text of the 2026 book, a second pass has to find the answer in that text before the question can ship, and a third pass that never sees the book tries to guess it. When Texas flipped on September 1, every Texas question that cited a renumbered article was re-verified against the 2026 text, and the ones that failed were regenerated. The same model family, gated, does not ship "Table 220.55." Ungated, it did.
What to do with this
Snapshot, not a verdict: six chatbots, one conversation, September 21, 2026, graded against the current bulletin and the 2026 code.
If you're studying for the Texas journeyman exam: buy the soft-bound 2026 NEC, tab it with the publisher's tabs, and write in it before test day. Use a chatbot to explain a concept or plan your weeks. When it gives you a section number, open the book and check that the section exists and says what the bot said. Right now, on the 2026 code, it usually won't. When it gives you a rule about the exam itself, read the bulletin; it's 25 pages and the part that matters is two.
If you'd rather practice on questions that were written from the 2026 text and checked against it, our Texas Journeyman NEC Knowledge practice exam and the Calculations portion are built the way described above. The first questions are free, so you can check our citations the same way we checked theirs.