GPT-4 and Claude 3 Opus Can Now Pass the CPA Exam With Flying Colors

Advanced artificial intelligence models have crossed a historic threshold in professional certification. As of late 2023 and throughout 2024, research indicates that top-tier large language models (LLMs), most notably OpenAI’s GPT-4 and Anthropic’s Claude 3 Opus, are now capable of passing all four sections of the Uniform Certified Public Accountant (CPA) Examination. This achievement marks a significant leap from earlier versions, such as GPT-3.5, which consistently failed the exam with scores hovering around 53%.

The current performance of AI on the CPA exam is not just a marginal success. In controlled studies, optimized versions of these models have achieved average scores of approximately 85.1%, significantly higher than the 75% passing threshold required by the American Institute of Certified Public Accountants (AICPA). This development is reshaping the understanding of cognitive automation in high-stakes professional services.

The Evolution of AI Performance on Accounting Certifications

The transition from AI as a "struggling student" to a "top-tier candidate" happened in less than a year. When ChatGPT (based on GPT-3.5) was first tested against accounting assessments in early 2023, the results were underwhelming. It struggled with complex mathematical logic, mixed up regulatory nuances, and frequently suffered from "hallucinations"—generating confident but incorrect accounting treatments.

From Failure to Mastery: GPT-3.5 vs. GPT-4

The initial testing of GPT-3.5 resulted in a failure rate that mirrored the most underprepared human candidates. The model found Financial Accounting and Reporting (FAR) particularly difficult, often failing to balance complex consolidation entries or misinterpreting the specific requirements of U.S. GAAP versus IFRS.

However, the release of GPT-4 introduced a paradigm shift. With increased parameter counts, better reasoning capabilities, and the ability to process much larger contexts, GPT-4’s performance skyrocketed. Researchers from institutions like Brigham Young University and the University of Duisburg-Essen found that while the base version of GPT-4 performed well, specific prompting techniques—such as "Chain-of-Thought" reasoning and "few-shot" learning—pushed the scores into the elite territory.

Comparing Top Models: GPT-4, Claude 3, and Gemini

The competitive landscape of AI means that GPT-4 is no longer the sole achiever in the accounting domain.

• GPT-4 (OpenAI): Remains the most consistent performer across all sections, especially in Regulation (REG) and Financial Accounting and Reporting (FAR). Its ability to handle structured tax law and standard accounting procedures is robust.
• Claude 3 Opus (Anthropic): Has demonstrated exceptional performance in the Auditing and Attestation (AUD) section. Claude’s architecture seems particularly suited for the nuanced, language-heavy scenarios found in audit reports and professional ethics questions.
• Google Gemini: Shows strength in specific data-heavy tasks but has historically lagged slightly behind GPT-4 in the strict logic required for complex multi-part accounting problems, though recent updates are closing this gap.

Breaking Down the Scores: Performance by CPA Section

The Uniform CPA Examination is currently structured around three "Core" sections and one "Discipline" section as part of the CPA Evolution initiative. AI performance varies across these domains based on the nature of the content.

Financial Accounting and Reporting (FAR)

FAR is traditionally considered the most difficult section for human candidates due to the sheer volume of material. Surprisingly, AI excels here when provided with clear numerical data. Since FAR relies heavily on the application of specific rules (U.S. GAAP), the AI’s ability to recall and apply these rules to Multiple Choice Questions (MCQs) is near-perfect.

Recent benchmarks show GPT-4 scoring above 80% in FAR. The model handles topics like revenue recognition (ASC 606), lease accounting (ASC 842), and government fund accounting with high precision, provided the prompts are structured to avoid calculation errors.

Auditing and Attestation (AUD)

AUD tests a candidate’s ability to apply professional skepticism and judgment. This was expected to be a weakness for AI, but the latest models have proven surprisingly adept at identifying audit risks and determining the correct audit opinion based on a given set of facts. Claude 3 Opus, in particular, has shown a refined ability to parse long, text-based narratives to find inconsistencies in internal controls.

Regulation (REG)

REG covers federal taxation and business law. This is a "lookup-intensive" section where AI has a natural advantage. Because tax codes are documented extensively in the training data, AI models can navigate the complexities of individual and entity taxation with ease. In many tests, AI performance in REG reached the high 80s, often surpassing human averages by a significant margin.

The New Disciplines: BAR, ISC, and TCP

Under the 2024 CPA Evolution, candidates choose one discipline.

• Business Analysis and Reporting (BAR): AI excels here due to the heavy emphasis on data analytics and financial management.
• Information Systems and Controls (ISC): This is a natural fit for AI, as it involves IT governance and cybersecurity—topics that are well-represented in AI training sets.
• Tax Compliance and Planning (TCP): Similar to REG, the AI's mastery of tax law makes this a high-scoring area.

How AI Conquered the Exam: The Technology of Reasoning

It is a misconception that AI passes the CPA exam through simple memorization. Because the exam is designed to test application and analysis, rote memorization of textbooks is insufficient. The AI uses several advanced cognitive strategies to solve these problems.

Few-Shot Prompting and Chain-of-Thought

Researchers found that providing the AI with a few examples of solved problems (few-shot prompting) significantly improved its performance. Furthermore, when the AI is instructed to "think step-by-step" (Chain-of-Thought), it breaks down complex accounting problems into logical sequences, reducing the likelihood of skipping a crucial consolidation step or a tax deduction limit.

Reasoning and Acting (ReAct) Frameworks

One of the breakthroughs in getting AI to pass the CPA exam was allowing the model to use external tools. In professional practice, an accountant uses a calculator, Excel, and tax research databases. When AI is given access to a "Python interpreter" or a "Calculator tool" during the exam simulation, its accuracy in the FAR and REG sections improves by nearly 9%. This allows the LLM to handle the linguistic interpretation of the question while delegating the precise math to a deterministic tool.

Handling Professional Judgment

The most impressive feat is the AI's burgeoning ability to simulate "professional judgment." By analyzing thousands of pages of AICPA audit standards and case studies, models can now weigh the "materiality" of an error or the "appropriateness" of an auditor's response in a way that mimics a seasoned professional.

Where AI Still Struggles: The Limits of Automation

Despite passing the exam, AI is not yet a "perfect" accountant. There are specific areas of the CPA exam where humans still maintain a competitive edge.

Task-Based Simulations (TBS)

Task-Based Simulations (TBS) account for 50% of the score in the Core sections. These require candidates to fill out tax forms, adjust journal entries in a spreadsheet format, or research specific code sections. While AI can handle the "logic" of these tasks, the complex, multi-tabbed interface of a real CPA exam simulation remains a challenge. AI often struggles with the synthesis of information across five or six different PDF exhibits provided in a single simulation.

The Hallucination Risk

"Hallucination" remains the Achilles' heel of LLMs. An AI might correctly identify the tax rule but then invent a specific dollar threshold that doesn't exist, or it might cite a non-existent GAAP sub-topic. In a professional setting, a 1% hallucination rate is a 100% liability risk. This is why AI is currently viewed as a high-powered assistant rather than a licensed practitioner.

Ethical Nuance and Real-World Skepticism

While AI can answer a multiple-choice question about ethics, it does not "experience" ethics. It lacks the real-world context of client pressure, firm culture, and the subtle cues of fraud that a human auditor might pick up during an in-person site visit.

Why This Matters for Current and Future CPAs

The fact that AI can pass the CPA exam is a signal that the entry-level tasks of the profession are being automated. This has profound implications for how accountants are trained and how they provide value.

The Shift to Advisory Services

As AI handles the "compliance" aspect of accounting—drafting financial statements, calculating tax liabilities, and checking audit boxes—the role of the CPA is shifting toward "Advisory." The accountant of the future will be a strategic consultant who interprets the AI-generated data to provide business insights, M&A advice, and long-term financial planning.

AI as the Ultimate Study Tool

For candidates currently preparing for the 2025 or 2026 CPA exams, the success of these models is good news. AI tutoring is becoming a viable and often superior alternative to traditional review courses.

• Instant Clarification: A candidate can ask an AI to explain the difference between a "finance lease" and an "operating lease" in five different ways until it clicks.
• Personalized Drilling: AI can analyze a candidate's practice test results and generate custom questions focusing specifically on their weak spots, such as governmental fund accounting.
• Simulation Preparation: Candidates can use AI to "roleplay" audit scenarios, helping them build the professional judgment required for the AUD section.

How to Use AI to Prepare for the CPA Exam

If you are a candidate looking to leverage these models, a strategic approach is necessary to ensure accuracy.

1. Verify with Primary Sources: Never take an AI’s word as the final law. Always cross-reference AI-generated tax thresholds with the latest AICPA blueprints or IRS publications.
2. Focus on Logic, Not Just Answers: Use the AI to explain the why behind a GAAP treatment. The goal is for you to pass the exam, not for the AI to do it for you.
3. Master the TBS Manually: Since AI struggles with complex simulations, you should spend more of your human study time on task-based simulations and use AI primarily for mastering the multiple-choice concepts.

Summary of AI vs. Human CPA Performance

The data is clear: AI has reached the level of professional competency required to pass the Uniform CPA Exam.

Conclusion

The question "Can AI pass the CPA exam?" has been answered with a definitive yes. The journey from GPT-3.5’s 53% to GPT-4’s 85% represents one of the fastest rates of professional skill acquisition in the history of technology. However, passing an exam is not the same as being a CPA. A CPA license carries legal weight, ethical responsibilities, and the requirement for human oversight that AI cannot fulfill.

For the accounting profession, this is not an existential threat but a technological evolution. The "CPA of the future" will not be someone who competes with AI to calculate numbers faster, but someone who masters AI to deliver deeper, more accurate, and more strategic value to their clients and organizations.

FAQ: AI and the CPA Exam

Which AI model is best for accounting?

Currently, GPT-4 and Claude 3 Opus are the top performers. GPT-4 is generally better for tax and financial reporting (REG and FAR), while Claude 3 Opus shows superior nuance in auditing (AUD).

Did ChatGPT pass the CPA exam on the first try?

No. The original ChatGPT (GPT-3.5) failed the exam. It was the successor, GPT-4, that passed after being given better reasoning prompts and access to computational tools.

Will AI replace accountants?

AI is unlikely to replace CPAs entirely. Instead, it will replace the repetitive, rules-based tasks within accounting. Licensed CPAs will still be required for signing audit reports, providing high-level tax strategy, and exercising professional judgment in complex legal environments.

Can I use AI to study for the 2025 CPA exam?

Yes, AI is an excellent study aid for explaining complex concepts and drilling multiple-choice questions. However, candidates must be wary of "hallucinations" and should always use AI in conjunction with official review materials like Becker, UWorld, or Gleim.

Why does AI struggle with task-based simulations (TBS)?

TBS requires synthesizing information from multiple disparate sources (like emails, invoices, and memos) and applying them to a specific document. AI currently has difficulty maintaining a perfect "mental map" of all these exhibits simultaneously, though this is expected to improve as context windows grow.