The JLPT has not published a vocabulary list since 2010, and most lists online derive from one reconstruction. We built our own from the evidence: 24,592 candidate words, judged twice by two AI models, with every decision and its reasons on the record.
Version 1.0N5 to N1, plus N1+CC BY-SA 4.0Unofficial
13,881words from N5 to N1
8,086more in N1+, useful beyond the exam
24,592candidates judged, each one twice
70.8%exact agreement between the two judges in the first round
Why we built it
One reconstruction became everyone’s list
Since the 2010 revision, the Japan Foundation and JEES have not published the vocabulary each JLPT level expects. The lists most learners use descend from the Tanos lists: one maintainer’s careful reconstruction of the old four-level lists, with N3 interpolated between them. Because nearly every app and deck copies Tanos, there was no independent check of where a word belongs.
Bebop teaches vocabulary one word at a time, so the level of each word matters. We wanted a list where you can see why every word is where it is, and disagree with a specific decision rather than with the whole list.
This is not an official list. The JLPT organisers did not write, review or endorse it. It is our estimate of the vocabulary expected at each level.
The result
13,881 words, larger than the old figures at every level
Each level came out 1.2 to 1.4 times the size of the pre-2010 specification. We tested what happens if a level is trimmed back to the old figure, and it removes words the official sample papers use (see the evidence), so the sizes are the judged levels, not a quota.
Cumulative words by levelThis listPre-2010 specification
N5969800+969
N42,0021,500+1,033
N34,6593,750+2,657
N28,6306,000+3,971
N113,88110,000+5,251
The right-hand column is the number of words added at each level. The old N3 figure is interpolated, because N3 did not exist before 2010. The old specification’s counting unit is not documented; this list counts each dictionary entry and reading separately and includes set phrases, counters and affixes, which gives larger numbers for the same vocabulary.
Method
Two judges, two rounds, then rules for what they couldn’t settle
Every candidate went through the same pipeline. In the first round neither judge saw the other’s answers, and the evidence did not include Bebop’s existing levels, which came from Tanos.
Gather candidatesEvery common word in the JMdict dictionary, every Tanos word, every content word in the official JLPT sample papers, and frequent words from three sentence corpora.24,592 candidates
Attach evidenceFrequency, sentence counts, appearances in the official samples by level, Tanos levels, kanji school grades and dictionary usage tags, for every word.8 evidence sources
Judge independentlyGPT-6 Astra and Claude each placed every word at a level or excluded it, in 246 batches of 100, with a short reason for each.68 calibration words in every batch
ReconsiderWhere they differed, each judge saw both answers and both reasons and could change its mind.7,171 disputes → round 2
Evidence ruleStill one level apart: the easier level if the word is in an official sample at that level, otherwise Tanos’s level if it is one of the two, otherwise the harder level.3,253 words
Third modelTwo or more levels apart, or included by only one judge: GPT-6 Sol, which gave neither answer, decided with the judges labelled only A and B.1,047 words
Human rulesPublished rules for place names and common-knowledge words, applied on top of the judged decisions, which stay on the record.77 words
Test and releaseLevel sizes tested against the official sample papers, cross-checked with the VDRJ research database, then released as version 1.0.v1.0, 26 September 2026
How each of the 24,592 candidates was decided
Agreed17,39970.8%Both judges gave the same answer in round 1.
Converged2,81611.5%The judges reconsidered each other’s reasons and reached the same answer in round 2.
Evidence rule3,25313.2%Still one level apart: the official sample papers, then Tanos, then the harder level.
Third model1,0474.3%Two or more levels apart, or included by only one judge: GPT-6 Sol decided.
Human rules770.3%Place names and common-knowledge words, decided by published rules.
The N1+ tier
N1 is for words an N1 examinee is realistically expected to know. Plenty of real, current words go beyond that, and excluding them would lose useful vocabulary. Bebop adds one tier, N1+, for them: 8,086 words. N1+ is not a JLPT level and has no size target.
Agreement
Where the two judges agreed
In the first round the judges gave the same answer for 70.8% of words, and were within one level of each other for 91.5%. Most disagreements were a single level, and in most of those GPT-6 Astra chose the easier one.
Midway through, the Claude judge changed from Claude Fable 5.1 to Claude Opus 5.5. On 500 words both judged, they agreed with each other 80.4% of the time, and each agreed with Astra about 69.5% of the time, so the switch did not move the Claude judge closer to or further from Astra.
91.5%of words within one level in round 1
3,631answers changed in round 2 (Astra 1,903, Claude 1,728)
Round-1 answers for all 24,592 candidates. The outlined diagonal is exact agreement; darker cells hold more words.
Examples
Every decision keeps its reasons
Each word in the list records how it was decided, both judges’ answers and their reasons. These five show each way a word can be decided.
食べるたべるto eat
GPT-6 AstraN5
Core beginner verb for daily routines, repeatedly attested in N5 samples.
ClaudeN5
Core beginner verb in every N5 syllabus and the N5 samples.
忖度そんたくsurmising someone’s wishes; deference
GPT-6 AstraN1 → N1+
Useful in political and workplace commentary, but prominence as a public discussion term does not establish core N1 expectation.
ClaudeN1+
Well-known political buzzword but formal and low frequency; beyond the N1 scale.
オリエンテーションおりえんてーしょんorientation
GPT-6 AstraN2
Standard school and workplace event vocabulary, suitable for N2 notices, schedules, and instructions.
ClaudeN1
School and company life word; Tanos N1.
雨雲あまぐもrain cloud
GPT-6 AstraN3
Ordinary weather vocabulary, including 雨雲レーダー forecasts; transparent meaning and everyday use support intermediate placement.
ClaudeN1+
Transparent weather word from known kanji; low frequency, not taught separately.
GPT-6 SolN2
An ordinary weather word, though less central than basic N3 weather vocabulary.
ヨーロッパよーろっぱEurope
GPT-6 AstraExcluded
Europe is a geographical region; the stated proper-name exception covers countries and languages, so it remains excluded.
ClaudeN3
Continent name used like country names in travel topics; Tatoeba 198, Tanos N3. Not a proper noun to exclude.
Evidence
Tested against the official sample papers
The official JLPT sample questions (the 2009 sample booklets and the 2012 official workbook, N5 to N1) are the closest thing to ground truth that is public. We counted which words occur at which level, and used them to test the list. Only the counts are kept; the text of the papers is not published.
1 in 30words of the N1 papers’ running text is missing from this list, against about 1 in 8 for Tanos
97%of the old specification’s words (as matched in the VDRJ database) are in the list
96%of VDRJ’s 5,000 most important words for students are at N5 to N1
Distinct words of each level’s papers coveredAs judgedTrimmed to the old figure
N393.4%86.3%
N294.5%82.3%
N196.6%91.6%
Trimming a level back to its pre-2010 size by frequency removes words that level’s own papers use. Frequency does not follow JLPT order: 鉛筆 (pencil) and 切手 (stamp) are N5 words near frequency rank 11,000. There is also no natural cut-off: each further thousand words adds two to three points of coverage. The axis starts at 75%.
Human rules
Where we stepped in, and why
After reviewing the exclusions, we added two rules. The judged decision stays on the record beside each curated one.
Place names
Countries, continents, seas and oceans are in, in their current names (ヨーロッパ at N4, 太平洋 at N2, キューバ at N1+). Cities, prefectures and regions stay out, except 東京 at N5. The judges had treated continents and less common countries inconsistently.
Common knowledge
General words the judges had excluded as proper nouns are in: オリンピック, 国連, the planets, era names such as 明治 and 令和, and Japanese history terms such as 藩 and 大名.
Sensitive words
Words that are vulgar in every sense are mostly excluded. Adult topics that appear in news and everyday use, such as 自殺, are included at their level and flagged, so apps built on the list can apply their own audience rules.
Limitations
What this list is not
The judges are language models. Their knowledge of what textbooks teach at each level is indirect, and two models can share the same errors. Agreement between them is evidence, not proof.
No qualified human editor has reviewed individual words yet. People decided the rules for disputes and the curation rules, not each word.
The official sample papers are small, about 900 to 2,000 sentences per level, so a word missing from them means little. Most are scanned, so their text comes from OCR.
Corpus counts are approximate, and the frequency data stops in 2021.
The list covers dictionary words. Grammar points and patterns are in a separate grammar list.
The size figures given to the judges came from pre-2010 guidance for a different test format. They steered the judges; the final sizes are the judged levels.
The list is licensed CC BY-SA 4.0. GPT-6 Astra and GPT-6 Sol ran through OpenAI’s Codex; the Claude judge was Claude Fable 5.1 and then Claude Opus 5.5.
In Bebop
Learn the whole list in your new tab
Bebop’s new tab teaches every word from N5 to N1: 13,901 illustrated cards, each with three example sentences. Choose any mix of levels. It’s free.