Work & Society · Nov 17, 2026
AI scores above 95 on six Japanese financial certification exams, but only 79.4 on the tax accountant exam
Testing AI on 321 questions from six exams, Sakana AI found top scores of 95.8 without tools and 97.3 with tools, from different models. On the tax accountant exam, the best score was 79.4
Koji Yamamoto · Economics Analyst

Key points
- On 321 questions from six Japanese financial certification exams, the top score was 95.8 without tools (Claude Fable 5) and 97.3 with tools (GPT-6 Astra)
- The questions came from six exams: the short-answer section of the CPA exam (93 questions), all 11 subjects of the certified tax accountant exam (40 questions), the Grade 1 Financial Planning Skills Test, the Grade 1 Financial Counter Services Skills Test, the Grade 1 JCCI Bookkeeping Test and the Grade 1 Zensho Bookkeeping Practice Test
- Scoring well on an exam is not the same as being useful on the job. Tax law changes every year, so people will still need to check the answers
AI models now score close to perfect on Japan's accounting and tax certification exams. Tokyo-based AI company Sakana AI tested current AI models on 321 questions drawn from Japanese financial certification exams. The top score was 95.8 without tools (Claude Fable 5) and 97.3 with access to a terminal and the web (GPT-6 Astra). The questions came from six exams in total: the short-answer section of the certified public accountant (CPA) exam, all 11 subjects of the certified tax accountant exam, the Grade 1 Financial Planning Skills Test and three others. Leading models scored between 96 and 100 on the five exams other than the tax accountant exam. On the tax accountant exam, the best score was 79.4.
For Japanese companies: Accounting and tax departments, and accounting firms, that plan to have AI draft journal entries or tax calculations should decide in December who will check the AI's answers and which documents they will check them against. December is when the busy season for year-end tax adjustments and annual tax returns begins. High exam scores do not change who is responsible for mistakes: the company and the licensed professional still are.
What was measured
Sakana AI named the benchmark "JFinBench" and published it on its website. It collects 321 questions from Japanese financial certification exams, has AI models answer them, and scores the results.
Two national accounting and tax qualifications make up 93 and 40 of the questions. The short-answer section of the CPA exam is the first stage of that exam. It is multiple choice and covers four subjects: financial accounting, management accounting, auditing and corporate law. The tax accountant exam has two accounting subjects, bookkeeping and financial statements, and nine tax law subjects: income tax law, corporate tax law, inheritance tax law, consumption tax law, liquor tax law, the National Tax Collection Act, inhabitant tax, enterprise tax and fixed asset tax. All 11 subjects are included in the question set.
These exams assume knowledge of Japanese law and accounting standards, and their questions turn on fine points of Japanese wording. Many have long assumed this put AI models, which learn largely from English-language data, at a built-in disadvantage.
95.8 without tools, 97.3 with tools
The top score was 95.8 without tools (Claude Fable 5) and 97.3 with tools (GPT-6 Astra). With tools, a model could use a terminal and the web to look up statutes and run calculations. Sakana AI said most failures came from gaps in knowledge of the Japanese system rather than from calculation errors. Even so, models were already scoring above 95 without tools.
The two top scores came from different models. Tools added 0.6 points for Claude Fable 5 and 3.7 points for GPT-6 Astra. On the tax accountant exam alone, the best score was 79.4. Finding out what accounts for the remaining lost points requires looking at each exam separately.
Exam scores are not the same as job performance
Readers should keep three caveats in mind when reading the scores.
First, each exam question has a single correct answer. In real accounting and tax work, documents are often incomplete and judgments can reasonably differ. An AI that scores 97 on an exam still cannot handle a case correctly if it does not know the company's circumstances.
Second, tax law changes every year. An AI does not know about amendments made after the point at which it was trained. It may confidently give answers based on outdated tax rates or deductions.
Third, the last few points matter. Getting three or four questions out of 100 wrong means that, on the job, a certain share of the output will contain errors. Errors in tax calculations can lead to additional tax assessments and late-payment penalties.
Without tools, GPT-6 Astra came second with 93.6, and Gemini 3.8 Flash came third with 88.8. The tools were a Linux terminal and the web. Models were not allowed to look at published answer keys, and any attempt that did received a score of 0. The questions came from six exams held between May and August 2026, and the authors checked all 321 questions against the original exam papers. Using the most recent exams, held in 2026, reduces the risk that models saw past questions during training.
What has changed
One common view held that language and institutional barriers protected professional work in Japan from AI. In these results, leading models scored between 96 and 100 on the five exams other than the tax accountant exam. At the same time, Sakana AI said the remaining failures came mainly from gaps in knowledge of the Japanese system.
The role of licensed professionals is what will change. Less of their work will involve producing answers from scratch. More will involve spotting where an AI's answer doesn't fit a company's circumstances or is out of step with the latest law, and correcting it. For accounting and tax departments, the question ahead is less about whether to use AI and more about how to set up a process for checking its answers.
Editorial cartoon
