Which AI chatbot gives the most accurate financial advice?

In Saturn's September 2026 study 'Artificial Authority: Should you trust AI to deliver financial advice?', the best performer was Claude Opus 5 in reasoning mode, which was still wrong 39% of the time. ChatGPT running GPT-5.6 Luna was wrong 58%, Grok 4.5 59%, Gemini 3.1 Pro 73%, and Claude Haiku 4.5 came last at 82%. Across all 18 models and more than 10,000 responses the average error rate was 57% — 49% for paid tiers, 63% for free tiers — rising to 88% on the hardest multi-step questions and 93% for free models on those. No model in the study reached the accuracy a regulated adviser is held to.

What did the AI chatbots get wrong about money?

Saturn counted an answer as a failure if it contained a factual error, omitted something material, or left out a warning that should have been there. Documented failures included a UK pension tax answer that would have exposed the user to a £17,500 HMRC charge; debt advice that prioritised clearing the highest-interest balance ahead of priority arrears such as council tax and rent, which can end in eviction or bailiffs; a fabricated rule letting graduates suspend student-loan repayments by emigrating (which would in fact raise their payments); calculation errors; and failure to account for scheduled tax changes. The questions covered debt, mortgages, pensions and tax, each asked five times to average out sampling variance.

Is the Saturn AI financial advice study reliable?

Treat the numbers as a vendor-commissioned benchmark, not an independent one. Saturn is a UK fintech that designed the methodology, applied the scoring and paid for the work, and it is lobbying the FCA to bring AI financial advice inside the regulatory perimeter — a position it would benefit from. That does not make the findings wrong: the question set covers 121 real consumer scenarios, each model was run five times, and the failure modes are specific and checkable. The FCA's own Mills Review found 26% of UK consumers already trust general-purpose AI with money questions, and a parallel PensionBee survey of 1,000 US adults found 57% would act on chatbot financial advice without independent verification and 23% had already received wrong information.

How should I use ChatGPT or Claude for personal finance safely?

Use them to understand concepts, draft questions for a human adviser, and check your own arithmetic — not to decide. Practical rules from the study's failure modes: pick a frontier reasoning model (Claude Opus 5 with extended thinking, GPT-6 Astra, Gemini 3.5 Pro) over a fast or free tier; state your jurisdiction and tax year explicitly; ask the model to list assumptions and cite the rule it is applying, then verify that rule on the official government or regulator site; never act on debt-prioritisation, pension-withdrawal or tax-election advice without a regulated adviser, since those are exactly where errors were most expensive; and remember that generic chatbot advice carries none of the redress protection that regulated advice does.

Quick Answer

ChatGPT vs Claude vs Gemini vs Grok: Financial Advice (2026)

Published: September 21, 2026

The headline numbers

UK fintech Saturn published Artificial Authority: Should you trust AI to deliver financial advice? in mid-September 2026, and the Financial Times ran it on September 20. The design: 18 AI models, 121 consumer money questions across debt, mortgages, pensions and tax, each asked five times, giving more than 10,000 responses. An answer failed if it contained a factual error, omitted something material, or missed a warning that should have been given.

57% of answers were wrong overall (43% accurate).
49% wrong for paid tiers; 63% for free tiers.
88% wrong on average for the hardest, multi-step questions; 93% for free models; the weakest models missed almost every complex query (up to 99%).

Verified September 21, 2026 against Saturn’s report coverage in the FT, Professional Adviser, IFA Magazine and ResultSense, and PensionBee’s 2026 AI and Your Money report.

Per-model results

Saturn named individual performers. Only the models the coverage identifies are listed; the remaining models in the 18 were not broken out publicly.

Model (as tested, Sep 2026)	Wrong answers	Notes
Claude Opus 5 (reasoning mode)	39%	Best in study; Anthropic’s default Opus since July 24, 2026
ChatGPT — GPT-5.6 Luna	58%	Consumer ChatGPT model tier at test time; GPT-6 Astra (Sep 3) not in the published breakdown
Grok 4.5	59%	xAI
Gemini 3.1 Pro	73%	Google; Gemini 3.5 Pro and 3.8 Flash not in the published breakdown
Claude Haiku 4.5	82%	Last place; a small, fast tier
Copilot and 13 others	not published	Included in the 57% average

Two readings follow. First, reasoning mode and model size matter more than brand: Anthropic fielded both the best and the worst model. Second, the spread between “best” and “worst” is 43 points, but even the best is wrong two times in five on questions an adviser must get right every time.

What the failures looked like

The specific errors are more instructive than the percentage:

Pension tax. One answer would have exposed the user to a £17,500 HMRC charge — the kind of allowance/threshold mistake a model makes when it applies last year’s rules or the wrong jurisdiction.
Debt prioritisation. Models steered users to clear the highest-interest balance first, ahead of priority arrears (council tax, rent). Mathematically tidy, legally dangerous: priority debts carry eviction and bailiff consequences that a credit card does not.
Fabricated rules. One model invented a provision letting graduates suspend student-loan repayments by emigrating — the opposite is true, and following it would have raised monthly payments.
Omitted risk warnings and failure to account for upcoming tax changes — both count as failures under Saturn’s rubric even when the arithmetic was right.

These map onto known LLM weaknesses: stale rules, jurisdiction blur, confident completion of a plausible-sounding rule, and optimising for the stated objective (minimise interest) rather than the unstated constraint (don’t get evicted).

The demand side: people are acting on it

PensionBee’s US survey of 1,000 adults who use chatbots for personal finance, released the same week, found 57% would act on the advice without independent verification, 23% had already received incorrect financial information from a chatbot, and 5% discovered the error only after acting. Gen Z and Millennials were the most willing to share sensitive financial data with a chatbot. In the UK, the FCA’s Mills Review put consumer trust in general-purpose AI for money questions at 26%, with the regulator warning that such advice carries none of the redress of regulated advice.

The provenance caveat

Saturn designed the test, scored it and paid for it, and it wants the FCA to regulate AI financial advice — a market it operates in. ResultSense’s summary is the right one: the figures are “the firm’s own and should be cited as such rather than treated as an independent benchmark.” The mitigating factor is reproducibility — 18 named models and a published question set make an independent rerun straightforward, and nobody without a commercial stake has done one yet.

How this fits the vendors’ own moves

Both OpenAI and Anthropic launched vertical finance products this month — ChatGPT for Financial Services and Claude for Financial Advisors — aimed at advisers and institutions, with retrieval over the firm’s own rules, compliance logging and human sign-off. The Saturn result is the argument for that architecture: a general chatbot answering a consumer cold is the worst configuration, a model grounded in current, jurisdiction-specific documents with a licensed human in the loop is what the products sell. The same logic runs through the Visa Trust Index on agentic commerce: consumers will delegate money decisions to agents faster than the agents earn it.

Using them anyway: a checklist

Model choice: frontier reasoning tier only — Claude Opus 5 or Fable 5.1 with extended thinking, GPT-6 Astra, Gemini 3.5 Pro. Never a free/fast tier for money.
Pin the facts: jurisdiction, tax year, residency, employment status, exact balances. Ask the model to restate them before answering.
Demand the rule: “Which specific rule or threshold are you applying, and where is it published?” Then open the government page.
Ask for the failure case: “What would make this advice wrong for me?” — this surfaces omitted warnings, the most common failure class.
Red lines: pension withdrawals, tax elections, debt prioritisation, mortgage products — get a regulated adviser. The study’s costliest errors were all in these categories.
Keep the transcript. If you do act on it, you will want the record; you will not have regulatory redress.