TreasuryBench 2026 report
How accurate is AI financial advice?
ChatGPT, Perplexity and four AI money apps answered the same 59 personal finance questions about two households. We scored all 354 answers.
By Junead Khan · Published · Updated · Tested September 2026
The short answer
In TreasuryBench's September 2026 test of 59 personal finance questions, general AI assistants gave better money answers than the AI built into budgeting apps. ChatGPT scored 89 out of 100 with no dangerous answers and Perplexity 84, against 61 for Quicken Simplifi and 53 for Monarch. Origin, an AI financial advisor app, scored 76. Across those five products, 10% of answers contained a wrong or out-of-date fact. Treasury, which runs the benchmark and makes the sixth product, scored 95.
01 / Overall results
General AI assistants averaged 86.5. AI budgeting apps averaged 57.
Six AI products answered the same 59 questions in September 2026. Each score is the average final task score out of 100, a blend of four quality dimensions, not the share of answers that were correct.
Treasury AI financial advisor app · AI Advisor
ChatGPT General AI assistant · GPT-6 Astra · Medium
Perplexity General AI assistant · Computer · Default Orchestrator
Origin AI financial advisor app · AI Advisor
Quicken Simplifi Budgeting app with built-in AI · Quicken AI Beta
Monarch Budgeting app with built-in AI · AI AssistantGeneral assistants: ChatGPT and Perplexity. Budgeting apps: Quicken Simplifi and Monarch. Compare prices and response times on the live leaderboard.
02 / Verdicts
What each AI is best and worst at with money.
Strongest and weakest of the 10 categories with three or more questions each.
- Treasury #1 of 6
95/100
AI financial advisor app
- Best at
- Employer Benefits & Perks 98
- Weakest at
- Retirement & Tax-Advantaged 91
- Dangerous answers
- 0 of 59
- ChatGPT #2 of 6
89/100
General AI assistant
- Best at
- Tax Strategy 96
- Weakest at
- Savings & Expense Reduction 80
- Dangerous answers
- 0 of 59
- Perplexity #3 of 6
84/100
General AI assistant
- Best at
- Insurance & Risk Protection 95
- Weakest at
- Credit Cards & Rewards 72
- Dangerous answers
- 2 of 59
- Origin #4 of 6
76/100
AI financial advisor app
- Best at
- Transaction Intelligence 85
- Weakest at
- Retirement & Tax-Advantaged 65
- Dangerous answers
- 0 of 59
- Quicken Simplifi #5 of 6
61/100
Budgeting app with built-in AI
- Best at
- Transaction Intelligence 88
- Weakest at
- Employer Benefits & Perks 37
- Dangerous answers
- 2 of 59
- Monarch #6 of 6
53/100
Budgeting app with built-in AI
- Best at
- Transaction Intelligence 75
- Weakest at
- Retirement & Tax-Advantaged 37
- Dangerous answers
- 5 of 59
03 / Your numbers
The budgeting apps that hold your data used it least.
Grounding scores whether an answer uses the household’s own balances, pay and bills. Monarch (44) and Quicken Simplifi (52) trailed ChatGPT (92) and Perplexity (90), though each product was set up for the same household through its own onboarding.
Used the household’s numbers and got the facts right · both ≥85
04 / One question, six answers
Same question, same household. Scores from 99 to 18.
“Am I contributing enough to my 401(k)?”
- Treasury 99
Found the match she is leaving behind and put a dollar figure on it.
“Microsoft's 401(k) plan matches 50% of every dollar you contribute up to the IRS elective deferral limit ($24,500 in 2026) … At your current pace, you will miss out on $9,850 in free employer money this year.”
- ChatGPT 94
Used her paycheck and the match; suggested raising her rate to 10%.
“Microsoft currently matches 50¢ per $1 you contribute, so at 8% you’d receive roughly another 4% ($2,400/year) from Microsoft … I’d increase your contribution from 8% to 10%.”
- Perplexity 94
Used her paycheck and the match; called her current rate reasonable.
“You're contributing 8% of your gross pay ($200 per semi-monthly paycheck …), which comes to $4,800/year … your current setup is in a reasonable range, not a red flag.”
- Origin 41
Read her contributions as $0 and fell back on general benchmarks.
“Your Origin profile currently records $0 in contributions for 2026 toward your Microsoft Corporation 401(k), leaving your full $24,500 elective deferral limit untouched for the year.”
- Monarch 35
Gave a checklist for her to work through instead of an answer.
“From your benefits or paystub, grab: Your contribution rate (like 5% pre‑tax, 3% Roth) … The match formula (like 100% of the first 4% of pay, or 50% of the first 6%).”
- Quicken Simplifi 18
Said it couldn’t tell from her data.
“I can’t tell from your current data whether you’re contributing enough to your 401(k).”
Excerpts from the published answers, trimmed with … for length and never reworded. Scores are each answer’s final task score. Read redacted excerpts of every answer.
05 / Safety
Nine answers could have cost someone money.
An answer is dangerous when its error could cause financial harm if followed. Treasury, ChatGPT, and Origin gave none.
- Treasury
0 dangerous · 2 with a factual error, of 59
- ChatGPT
0 dangerous · 1 with a factual error, of 59
- Perplexity
2 dangerous · 8 with a factual error, of 59
- Origin
0 dangerous · 8 with a factual error, of 59
- Quicken Simplifi
2 dangerous · 4 with a factual error, of 59
- Monarch
5 dangerous · 9 with a factual error, of 59
06 / Every category
Scores by topic, from budgeting to tax and retirement.
Average task score out of 100. Bold marks each category’s top score, ties included. Pilot categories have fewer than three questions, so read them as early signals.
| Category | Treasury | ChatGPT | Perplexity | Origin | Quicken Simplifi | Monarch |
|---|---|---|---|---|---|---|
| Debt & Credit Health2 questionspilot | 99 | 99 | 86 | 95 | 93 | 88 |
| Employer Benefits & Perks4 questions | 98 | 89 | 73 | 82 | 37 | 44 |
| Housing & Rent4 questions | 97 | 93 | 90 | 79 | 57 | 56 |
| Transaction Intelligence8 questions | 97 | 91 | 89 | 85 | 88 | 75 |
| Investing & Equity Comp4 questions | 96 | 90 | 86 | 67 | 49 | 56 |
| Insurance & Risk Protection4 questions | 96 | 91 | 95 | 84 | 60 | 67 |
| Credit Cards & Rewards8 questions | 95 | 82 | 72 | 74 | 50 | 38 |
| Life Planning & Major Decisions2 questionspilot | 94 | 95 | 81 | 75 | 62 | 50 |
| Cashflow & Budgeting5 questions | 93 | 87 | 84 | 73 | 60 | 63 |
| Savings & Expense Reduction4 questions | 93 | 80 | 74 | 72 | 62 | 51 |
| Tax Strategy8 questions | 93 | 96 | 88 | 73 | 57 | 43 |
| Retirement & Tax-Advantaged6 questions | 91 | 88 | 91 | 65 | 60 | 37 |
| Overall | 95 | 89 | 84 | 76 | 61 | 53 |
Columns: Treasury, ChatGPT, Perplexity, Origin, Quicken Simplifi, Monarch.
07 / What to do
How to use AI for financial advice safely.
Give it your real numbers
Answers that used the household’s balances and spending scored highest. Generic answers scored lowest.
Check rule-based figures
Verify contribution limits, tax brackets and benefit terms against the official source.
Slow down on the hard topics
Credit-card rewards, employer benefits and retirement had the widest spread between products.
Treat one answer as a start
Our guide shows how to check AI financial advice before you act on it. Read the guide
08 / Method
How TreasuryBench tested the six products.
- Step 1
Set up
Each product was set up through its own onboarding for the same two synthetic households: a single renter in Seattle and a family with a toddler in Denver.
- Step 2
Ask
Each was asked the same 59 plain questions across 12 categories. Many hide an opportunity the household’s data points to.
- Step 3
Score
Every answer was scored on grounding, correctness, resolution and prudence. Automated checks are blended with an AI judge. Errors cap the score: 40 for a dangerous contradiction of a known fact, 55 for several material contradictions, 65 for one, and 70 to 80 for stale or wrong current facts. An incomplete or cut-off answer is capped at 85.
The full rules are in the scoring rubric.
Limits and disclosure
- Two synthetic households and 59 questions cannot cover every situation. This is not a statistically significant ranking.
- Published answers are redacted excerpts, not complete verbatim captures: account identifiers, subscription and budget details, profile references and data-source labels were removed.
- The release publishes redacted answer excerpts and scores, not setup logs or account configurations, so it does not show that every product took in the data identically.
- Each product was tested once, in the version shown, during September 2026. Products change, and a later run may differ.
- An AI judge can make mistakes. Redacted answer excerpts and per-question scores are public so anyone can check them; the full inputs behind each judgment are not.
- Treasury designed the benchmark, scored the answers and makes one of the products.
09 / FAQ
Questions people ask.
Is ChatGPT good for personal finance questions?
In TreasuryBench's September 2026 test, ChatGPT (GPT-6 Astra · Medium) scored 89 out of 100 on 59 personal-finance questions, the highest of the products Treasury does not make. One of its 59 answers contained a factual error and none was dangerous. It scored highest of all six products on tax strategy. Check any figure that depends on a current rule, such as a contribution limit, against the official source.
What is the best AI for personal finance questions?
In the September 2026 TreasuryBench run the scores out of 100 were Treasury 95, ChatGPT 89, Perplexity 84, Origin 76, Quicken Simplifi 61, and Monarch 53. Among the products Treasury does not make, ChatGPT scored highest. Treasury runs the benchmark and makes one of the products, so the questions, redacted answer excerpts and scores are public for anyone to check.
Are AI budgeting apps like Monarch or Quicken Simplifi accurate?
In TreasuryBench's September 2026 test their AI assistants trailed general AI assistants. Monarch's assistant scored 53 out of 100 and Quicken Simplifi's 61. Monarch had 9 answers with a factual error and 5 dangerous ones; Simplifi had 4 and 2. Both trailed most on using the household's own numbers (grounding 44 and 52, against 92 for ChatGPT).
How accurate is AI financial advice?
It depends on the product and the question. In TreasuryBench's September 2026 test, across the five products Treasury does not make, 30 of 295 answers (10%) contained a wrong or out-of-date fact and 9 could have cost money. TreasuryBench scores are composite quality scores out of 100, not the percentage of answers that were correct.
Who runs TreasuryBench, and can I check the results?
Treasury runs it and makes one of the six products tested. The questions, redacted excerpts of every product's final answers, per-question scores and the scoring rubric are public on GitHub, and aggregate scores are a CSV download. Treasury designed the questions and scored the answers.
Cite and reuse
Cite this report.
Khan, Junead (Treasury). "TreasuryBench 2026: How accurate is AI financial advice?" Published October 17, 2026. https://treasury.sh/benchmarks/2026. Data: https://github.com/Treasury-Technologies-Inc/treasurybench/tree/main/results/2026-09.
Include the test date (September 2026) when you quote a score. Questions, redacted answer excerpts and scores · Scoring rubric · Download scores (CSV) · GitHub repository
