Can ChatGPT answer tax questions? GPT-6 Astra got 80.8% in our test

Yes for simple questions: 11 AI models on 26 tax questions

Whizi's Canadian Tax LLM Bench asked 11 models the same 26 questions on the 2025 Canadian tax year, two for each province and territory. OpenAI's GPT-6 Astra got 80.8% right. Eight of the 11 scored under a third.

We worked every key from the Canada Revenue Agency's or the province's published rates. Then a second agent worked it again without seeing the first answer. Each model answered once, at low reasoning, with the same short system prompt and no web search. A full return had to land within one dollar of the key, and a receipt had to match to the cent. Close counted as wrong. We tested the models themselves, not the ChatGPT app, which can search the web and may answer differently.

That's a small test. With 26 questions, GPT-6 Astra's 80.8% carries a 95% confidence range of plus or minus 14.7 points, and Claude Opus 5.5's 53.8% carries 17.9. The two ranges overlap, so the gap could be smaller than it looks. The questions are Canadian, too. We haven't run a US version, so read the scores as a measure of how models handle tax arithmetic, not as a grade for your US return.

Where do AI models get tax questions wrong?

AI models get tax questions wrong on whole returns, the questions that put several kinds of income, deductions and credits on one return. The bench has 14 of them. Seven of the 11 models got none right, and only GPT-6 Astra passed half.

The easy column tells the same story from the other side. Benefits and basics holds anchors such as the 2025 RRSP limit, and every model scored above zero there. Three got all five. The column counts are small (two questions for sales tax, two for a corporation), so one answer moves a cell by 50 points. The same table sits in our AI in accounting statistics.

A retired couple in British Columbia shows how close-but-wrong happens. The key for their combined 2025 tax, with half of one spouse's RRIF withdrawals split to the other, is $26,065. GPT-6 Astra matched it. Claude Opus 5.5 answered $26,051 and Claude Fable 5.1 answered $26,069, and both scored wrong.

Opus showed its work and named its own assumptions, starting with the federal rate. That rate is a trap in 2025. The CRA says "The federal tax rate decreased from 15% to 14% on July 1, 2025, making the tax rate 14.5% for 2025." Opus used 14.5% and still missed by $14.

A receipt from Calgary tests a date. The shopper buys a restaurant meal and a printed novel on 15 February 2025, then a child's jacket and an adult's winter coat on 16 February. Canada's GST/HST break ran "from December 14, 2024, to February 15, 2025", per the federal government's page, so only the second day's purchases carried GST. The key is $15.50. Gemini 3.8 Flash, Claude Sonnet 5 and GLM-5.3 answered $21.75, which taxes all four items as if the break never happened.

Some failures aren't wrong answers at all. DeepSeek V4.1 Flash returned nothing on 19 of the 26 questions. It hit its token cap before writing a word, and a blank scores wrong. You can read every public question and each model's answer on the bench questions page.

Is ChatGPT accurate for taxes in the US?

ChatGPT was right on 39 to 47 percent of common tax questions in the one peer-reviewed test we found, by two accounting researchers at US universities. Hannah Smith Antinozzi and Lauren Cooper entered the questions into ChatGPT 3.5 and 4 during the 2023 and 2024 tax seasons. The Journal of Emerging Technologies in Accounting published the paper online on 21 March 2025.

They found lower accuracy on questions that are more common, that have more complex answers, that "require an evaluation of taxpayers' fact patterns", and that cover rules set after the model's knowledge cutoff. They found no difference between the two versions. Their conclusion: ChatGPT "is generally not a reliable source of tax guidance for uninformed taxpayers."

Those models are old by October 2026, and newer isn't automatically better. On our Canadian set, a different test, only GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 beat the study's 47 percent. The pattern carries over, though. In both tests, the questions that need a taxpayer's own facts scored lower than the rest. In our bench, every one of the 11 models scored lower on whole returns than overall.

Can AI do my taxes for me?

AI can draft parts of a return, and you're still the one who answers for it. The IRS Taxpayer Advocate Service puts the rule plainly: "You are legally responsible for everything listed on your tax return even if you followed the advice of a tax return preparer." A chatbot isn't even a preparer.

The bench shows which part goes wrong most. A whole return, with salary, dividends, a capital gain and credits on one page, is where 7 of 11 models scored zero. A question like the 2025 RRSP limit is the kind they mostly got right. That's a description of the test, not a rule for your return.

Before you paste anything into a chatbot, replace your name, address and Social Security number or SIN with placeholders. A tax question never needs them. Our guide to uploading documents to AI covers what each vendor keeps.

When should you check with a CPA, the IRS or the CRA?

Check with a person or the tax authority whenever the answer depends on your own numbers. In the bench, all 11 models scored lower on whole returns, the questions built on one household's full set of figures, than on the test overall. Both governments answer questions for free, as of October 2026.

The CRA's free clinics turn away the same situations the models handled worst. Self-employment income, business income, rental income, capital gains and interest over $1,200 are all on the clinics' not-eligible list, per the CRA page modified 26 October 2026. On the bench's three self-employed questions, no model but GPT-6 Astra got more than one right. For those returns, the CRA line, a CPA or another tax professional is the route.

For anything that moves real money, from a business return to a sale of property, a CPA can look at your actual documents. No model in our test did that. Each one answered from memory, without seeing a single slip.

How to ask ChatGPT a tax question you can check

Ask ChatGPT for the rule and where it lives, not for your number. Paste the prompt below with the brackets filled in.

I have a question about [US federal / Canadian federal and my province] tax for the [year] tax year: [question, with no names or ID numbers]. Explain the rule in plain English. Name the official form and line, or the IRS or CRA page, where it is set out. Say which facts about my situation would change the answer. Do not calculate what I owe. End with the questions I should ask a tax professional.

"Name the official form and line" gives you something to look up on irs.gov or canada.ca. "Do not calculate what I owe" keeps the model off the whole-return arithmetic where the bench scored lowest. The last line turns the answer into a list for a preparer.

Then ask a second model the same question. In the British Columbia example, three models gave three totals within $18 of each other, and only one matched the key. Two models agreeing isn't proof. A disagreement is your cue to open the official page.

In Whizi, side by side model chats (Pro and Powerhouse) send one question to two models in two columns, and the compare guide shows how. GPT-6 Astra, the bench's top scorer, is on Powerhouse only. Claude Opus 5.5 is on Pro. Whizi's Canadian CPA agent searches only official Canadian sites such as canada.ca. The bench didn't test that agent, and its base model, GPT-6 Luna, scored 19.2% at low reasoning with search off. Where this loses: the IRS assistant, both phone lines and the free clinics cost nothing and come from the tax authority itself.