Finance

ChatGPT, Claude and Gemini on Austrian tax returns: four test cases, one clear verdict

GPT-5, Claude Opus 4.7 and Gemini 2.5 all claim fluency in Austrian tax law. Four real-world cases show where they deliver — and where they cost users thousands of euros.

Option News Redaktion · 6. Juni 2026 · 11 Min. Lesezeit

Laptop showing a ChatGPT interface next to a printed FinanzOnline form on a Viennese desk

Anyone preparing an Austrian tax return in 2026 will have toyed with the same thought experiment: GPT-5 has been available since February, Claude Opus 4.7 since March, Gemini 2.5 Pro since November 2025. All three claim fluency in German-language tax law. And all three are now quietly used by freelancers, ETF savers and private investors to structure the Anlage E1, the Beilage K1 or the list of deemed distributions on accumulating funds.

The question is not whether this works. The question is when it works — and when a language model walks the user, with full confidence, into a four- or five-figure tax mistake. We ran four representative Austrian cases through each of the three leading models and compared the answers against the actual law. The picture is considerably more mixed than the vendors' polished demos suggest.

The test parameters, up front:

  • Four constructed cases, all distilled from real client queries handled by eastern Austrian tax advisers
  • Three models in their publicly available consumer tier (ChatGPT Plus, claude.ai Pro, Gemini Advanced)
  • Identical German-language prompts, no plug-ins, no web browsing, no retrieval augmentation
  • Cross-checked against FinanzOnline, BMF guidance and the income tax code as of June 2026

Why taxpayers turn to language models at all

A Viennese tax adviser costs between EUR 150 and 350 an hour in 2026. A sole trader with EUR 60,000 to 90,000 in turnover rarely pays less than EUR 900 a year for a complete income- and VAT-return package. For private investors with a plain-vanilla ETF and crypto portfolio, that figure quickly looks like a meaningful slice of their realised gains — a line item to cut.

That is precisely the group that has been testing, for months now, whether a well-prompted language model can deliver the first draft of a return. The hope is not to replace full advice, but to structure one's own records: which receipt belongs to which line on the form, how a freelance invoice should be handled for VAT, when a crypto swap counts as a taxable realisation event.

That hope is not unreasonable. It is, however, asymmetric in its risk. A model that quotes the correct flat rate in 90% of cases and hallucinates the wrong percentage in the remaining 10% — with the same confident tone — is more dangerous than a model that admits "I don't know". Our four test cases confirmed exactly that.

Case 1: Self-employed copywriter, freelance contract, EUR 80,000 gross

Anna K., 34, works on freelance contracts for two Viennese agencies. In 2025 she invoiced EUR 80,000 net. She uses the basic flat-rate deduction for business expenses (12%) and pays around EUR 18,500 in social-insurance (SVS) contributions. She is above Austria's 2025 small-business VAT threshold of EUR 55,000 and therefore subject to standard VAT.

The prompt to all three models: "I am a freelance copywriter in Vienna, 2025 net turnover EUR 80,000, basic flat-rate deduction, SVS contributions around EUR 18,500, no dedicated office. Which annexes do I need for my income tax return and roughly what tax bill should I expect?"

GPT-5 produced a cleanly structured answer: Anlage E1 plus E1a, correct mention of SVS deductibility, a plausible rough estimate of EUR 19,500 in income tax. The model did, however, state the small-business VAT threshold as EUR 35,000 rather than the EUR 55,000 in force since 2025 — a figure that did not affect this particular answer, but in a neighbouring scenario would have triggered the wrong "you qualify as a small business" recommendation.

Claude Opus 4.7 recognised the freelance setup, actively asked for additional details (commuter allowance, special expenses, church contributions) and added that the basic flat-rate deduction for liberal professions falls under line 9259. The line reference is correct; its estimate of EUR 19,700 sits close to the actual liability of EUR 19,620.

Gemini 2.5 also produced the correct combination of annexes but referenced the long-defunct Solidaritätszuschlag, a surcharge that has never existed in Austria — a classic translation slip from German training material. Its final figure of EUR 17,400 came in about EUR 2,200 light.

Case 2: EUR 12,000 of crypto gains across mixed positions

Mehmet S., 41, has held Bitcoin and Ethereum in a Bitpanda wallet since 2019. During 2025 he made several partial sales for a combined gain of EUR 12,000. Half of the coins sold were acquired after 1 March 2022 (after Austria's crypto-tax regime under the Ökosoziales Steuerreformgesetz took effect); the other half before.

The clean legal answer: coins acquired before 28 February 2021 qualify as legacy holdings (Altbestand) and are tax-free. Coins bought from 1 March 2021 onwards are subject to the 27.5% KESt. Realisation follows FIFO unless a clear lot-by-lot allocation is documented. Bitpanda has reported KESt automatically to the BMF since March 2024 — the investor still has to check that all Altbestand confirmations are on file.

GPT-5 recognised the legacy rule and named the cut-off date correctly. It missed, however, that crypto-to-crypto swaps have not counted as a realisation event since 1 March 2022 — a point not strictly relevant to Mehmet's case, but one that in a slightly different setup would have led to double taxation.

Claude Opus 4.7 first asked for a complete transaction history, then delivered a correct FIFO calculation and explicitly flagged the Bitpanda automation. It was the only one of the three models that unambiguously recommended pulling the broker's annual tax certificate and reconciling it against the user's own records.

Gemini 2.5 quoted the wrong rate at 25% instead of 27.5% — a transfer from the German Abgeltungsteuer. On EUR 12,000 of gains that is a EUR 300 gap. In larger portfolios it scales accordingly.

Case 3: The ETF savings plan that turns into a trap

Lisa T., 29, has put EUR 250 a month since 2022 into an accumulating, US-domiciled Vanguard ETF. She bought the product through a foreign online broker that does not offer Austrian automatic KESt handling. The ETF is not reported to the OeKB.

This is the most dangerous case in the test, because at first glance it looks like an ordinary ETF savings plan — yet it is taxed in a completely different bucket from the standard Austrian setup. An unreported foreign ETF is subject to the flat-rate income regime under section 186 of the Austrian Investment Fund Act: 90% of the calendar-year price appreciation, but at least 10% of the year-end value, counts as a deemed distribution and is taxed at 27.5%.

We described the full setup to each model and asked: "How do I declare this ETF on Anlage E1kv?"

GPT-5 recognised the facts and quoted the punitive estimate rule correctly. It failed to deliver the crucial practical recommendation: that the investor should liquidate the position promptly and switch into an OeKB-reported UCITS equivalent — the advice every Austrian tax adviser would give at this point.

Claude Opus 4.7 walked through the legally correct calculation and added the practical advice to switch into a reported variant. It also noted that the punitive estimate is typically higher than the actual appreciation, which makes the switch worthwhile even when crystallising a loss.

Gemini 2.5 failed this case outright. The model treated the ETF as if it were a German Vorabpauschale product, quoted the hypothetical reference rate from the German Investment Tax Act and produced a tax figure roughly a third of the correct one. Any taxpayer following that answer is heading for a guaranteed amended assessment with interest charges.

Case 4: Commuter allowance, Tulln to Vienna

Thomas R., 47, commutes daily from Tulln to Vienna-Heiligenstadt. The one-way distance is 32 km; the trip is feasible by public transport in under 90 minutes. His annual gross salary is EUR 62,000.

The correct answer: because public transport is deemed reasonable, the small Pendlerpauschale applies. For a distance between 20 and 40 km it amounts to EUR 696 per year. He is also entitled to the Pendlereuro at EUR 2 per kilometre of one-way distance per year (EUR 64). The Klimaticket Wien-NÖ-Burgenland, if he holds one, is deductible as work-related expenses — but cancels the Pendlerpauschale outright if the employer covers the entire cost of the ticket.

GPT-5 produced the correct figures for both the allowance and the Pendlereuro. It missed the offset question on the Klimaticket: if the employer pays for the ticket, the allowance falls away; if not, both stand. The distinction made no difference in Thomas's actual case because he pays for the ticket himself, but in a variant with an employer-funded Klimaticket the model would have recommended double-counting.

Claude Opus 4.7 actively asked about employer involvement in the Klimaticket, presented both scenarios in parallel and added the correct line number 718 for the employee assessment form. Fully accurate.

Gemini 2.5 quoted an outdated 2022 allowance of EUR 372 and omitted the Pendlereuro entirely. A gap of EUR 388 per year.

The scorecard at a glance

| Case | Correct answer | GPT-5 | Claude 4.7 | Gemini 2.5 | Serious error? | |---|---|---|---|---|---| | Freelance EUR 80k | Income tax ~EUR 19,620, E1+E1a | EUR 19,500, correct | EUR 19,700, correct | EUR 17,400 | Gemini: EUR -2,200 | | Crypto EUR 12k gain | KESt 27.5% on post-cutoff lots | correct, swap rule missed | fully correct | 25% instead of 27.5% | Gemini: wrong rate | | US ETF, not OeKB-reported | Punitive 90/10 estimate | correct, switch advice missing | fully correct | German Vorabpauschale | Gemini: off by a factor of three | | Commuter allowance 32 km | EUR 696 + EUR 64 Pendlereuro | correct, Klimaticket gap | fully correct | EUR 372, no Pendlereuro | Gemini: EUR -388 |

Who benefits from the tool — and who does not

The honest verdict after four cases: language models are a useful sparring partner for structurally simple situations in which the user can broadly sanity-check the answer. A sole trader who has invoiced the same freelance contracts all year can have Claude or GPT-5 walk through the annexes and perhaps save EUR 200 in adviser fees.

Anyone holding foreign brokerage accounts, swapping crypto, contributing to a private foundation, drawing foreign income or changing legal form mid-year is not in the target group. The gap between "90% correct" and "100% correct" is, in tax matters, identical to the risk of an amended assessment — and the latter, with surcharges, swiftly costs more than a decade of professional fees.

A Viennese tax adviser we discussed the test with put it, paraphrased, like this: he uses these tools himself for first-pass structuring and client letters, has no fear of them, but considers it a question of professional indemnity that the end client should never bear sole responsibility for an answer produced by a model. Savings made there will be wiped out by the first serious slip.

Alternatives with a clear scope

Three tools now play in the same league as the generic language models in 2026, but each covers a tightly defined segment.

The FinanzOnline assistant from the BMF has been available for employee tax assessments since February 2026. It walks the user through the standard prompts (special expenses, extraordinary burdens, commuter allowance) and writes the answers directly into the return. It is not suited for the self-employed or investors; for pure salary cases it is the superior tool — free, legally watertight, no model risk.

taxefy.com is an Austrian tax-software start-up that has offered guided employee assessments since 2023 and, since early 2026, can import crypto income directly from Bitpanda and Coinbase. For investors with a manageable portfolio it is the clean solution; the self-employed get less out of it.

A traditional tax adviser remains unavoidable for complex situations. A scope of two to three hours a year (typical for a sole trader on EUR 80,000-150,000) costs EUR 400 to 1,000 and buys, alongside the return itself, a reversal of the burden of proof in the event of an audit. That insurance value is something no language model can replicate.

For a deeper dive into the Austrian treatment of capital income, see the full guide to buying ETFs in Austria and the outlook on the planned KESt reform in Austria's 2027 KESt overhaul.