We asked ChatGPT if a couple can retire at 55 with $2 million. Here's what it missed.

Plan Mirror · Updated 2026-11-07

Ask ChatGPT whether you can afford to retire and you'll get a confident, well-written answer in under a minute. We wanted to know what that answer is worth. So we gave it a real-looking question, a Texas couple retiring at 55 with $2 million, answered every follow-up it asked, and then ran the same household through Plan Mirror.

The short version: ChatGPT spent five replies working the answer out from scratch, revising it as each fact arrived, and finished without giving the couple the one thing they asked for, an answer they could rely on. Plan Mirror gave a clear one. The plan has a hole in it at 59. ChatGPT saw the hole and suggested fixes, but it couldn't tell the couple which of them would work. Plan Mirror tested each one: switching to 60/40, ChatGPT's headline advice, doesn't help, and a small 72(t) schedule, which it offered only as a backup, does the most.

The question

We sent one message, the kind anyone might type:

My spouse and I are both 54 and plan to retire in January 2027, when we'll both be 55. We live in Texas. We've saved about $2 million: $1.6 million in traditional 401(k)s and IRAs, $300,000 in a regular taxable brokerage account, and $100,000 in Roth IRAs. We expect to spend about $90,000 a year. Our combined Social Security will be about $48,000 a year if we both start at 67. We want the money to last until we're 95. Can we afford to retire in 2027, and how should we draw down our accounts?

Then we answered ChatGPT's questions, one per reply: the $90,000 includes health insurance and taxes, all the pre-tax money is in IRAs rolled over from old jobs, the Roth is all original contributions, and everything is in stock index funds.

What a ChatGPT session gets you: a model built on the fly

ChatGPT doesn't have a retirement model. It writes one in the conversation, as Python code, and rewrites it as it goes. That's impressive to watch. It's also the problem.

The first model was a sketch. Its opening calculation assumed “Constant inflation-adjusted portfolio returns.” and “No separate deductions for income taxes, investment fees, or additional health care expenses.” No market crashes, no taxes. On that basis it called retiring in 2027 “a cautiously favorable plan rather than a guaranteed one”.

It assumed before it asked. Our message said "401(k)s and IRAs", and ChatGPT ran with the best case, telling the couple “you may qualify for the IRS separation-from-service exception, often called the Rule of 55.” Two replies later, once it asked and learned the money was all in IRAs: “That changes my recommendation in an important way.”

The numbers moved mid-conversation. Adding up the money the couple can reach before 59½, it found “Only a $5,000 difference before investment returns, inflation, and cash-flow timing.” A few paragraphs later it remembered inflation: “You could need roughly $430,000 over the bridge period”.

The verdict kept changing. "Cautiously favorable" became “I would lean toward retiring as planned in January 2027”, then “feasible with meaningful market risk—not yet proven safe”, then, when it learned they were all in stocks, “This changes my retirement recommendation more than anything else you've told me.”

It found the gap but couldn't size it. ChatGPT spotted the problem this page is about: the years before 59½, when most of the money is locked in IRAs. In one reply the bridge was “plausible”; in the next, “does not fully close the bridge”. It never said how likely the bridge was to fail.

And after five replies, no odds. Its own words: “We haven't yet established a numerical probability of success.”

None of this is ChatGPT being careless. It's what happens when the model is invented fresh for each conversation. Ask again tomorrow and you may get different code, different assumptions and possibly a different answer. There's no engine behind it to check the numbers against, and at no point did the conversation produce a full simulation with taxes, market swings and the rules on when each account can be touched.

What Plan Mirror gives you: one engine, five tests

Plan Mirror doesn't write a new model for each question. The same calculation engine runs every plan, and the same inputs always give the same results. It's covered by an automated test suite, and every scenario we publish is run on it, so it gets exercised on a wide range of households. You can read how it works on the methodology page.

Here's the same couple, with the same facts we gave ChatGPT:

Household

A couple, both 55

Retires

At 55, in 2027

Plans through

Age 95 (2067)

Invested

$2,000,000: $300,000 taxable brokerage, $1,600,000 traditional 401(k)/IRA, $100,000 Roth

Spends

$90,000 a year in today's dollars, taxes and healthcare included

Social Security

$48,000 a year combined, in today's dollars, from 67

401(k)/IRA before 59½

Not touched until 59½

Assumes

6% average return, 2.5% inflation

Taxes

Married filing jointly, Texas (2026 rules)

Spending is all-in, as we told ChatGPT. The $2 million is today's balance; Plan Mirror grows it to $2.1 million by the day they retire, and inflation makes the first year cost $92,250. By default Plan Mirror keeps the IRAs closed until the calendar year they turn 60, spending the brokerage account first and the Roth contributions next.

Test one: steady returns

Plan Mirror grows the portfolio at 6% a year with 2.5% inflation and works out every year's taxes. Over the long run the couple is fine: they finish at 95 with $2.6 million in today's dollars.

But look at 2031. Spending has grown to $101,827, the brokerage account is gone and the Roth can cover only $92,005. The couple is sitting on $2.3 million, nearly all of it in IRAs they can't touch yet. That's a shortfall at 59, with millions in the bank.

Dollar figures in the table are in each year's own dollars.

Tests two and three: the stress tests

With spending 10% higher they still end with $1.9 million. With returns 2 points lower, $591,000. Both hit the same wall in 2031. The long run has plenty of room. The bridge has none.

Test four: a century of real markets

Plan Mirror replays the plan through actual U.S. stock returns and inflation from every start date since 1926 Q3.

152 of 237 start dates make it: 64.1%. Retiring into the Depression (1929 Q4) runs them short by 2029, at 57. And every failing history fails before Social Security starts: 35.9% by then, 35.9% by 95. In none of them does the couple run out of money in old age. They run out of money they can reach.

Test five: 1,000 possible futures

Plan Mirror generates 1,000 random market paths around the same average return, swinging at an assumed 15% volatility a year.

477 of 1,000 succeed: 47.7%. Again it's early: 50.4% of paths run short before Social Security, out of 52.3% in total.

That's the number ChatGPT couldn't give. The plan works in roughly two in three historical starts and one in two simulated futures, and nearly all the failures sit in the five years before the IRAs open. These are shares of tested market paths, not a guarantee either way, but they tell you where the risk is.

Does 60/40 help?

ChatGPT's headline recommendation was to get out of all-stocks: “I would use 60% stocks / 40% bonds and cash as an initial planning target”. Plan Mirror can test a 60/40 mix in seconds. With 40% in bonds across the portfolio:

Worse on history, 56.1% against 64.1%, and no better on random futures, 47.3% against 47.7%. Bonds don't fix a plan whose problem is reaching the money, not losing it. Advice that sounds right still needs a test before you act on it.

The insurance surprise

One more thing ChatGPT missed. It warned the couple about the ACA subsidy cliff, where the premium tax credit ends above 400% of the poverty line. It didn't mention the floor. While they live on the brokerage account and the Roth, their counted income is tiny, 85.1% of the poverty line in 2027, and below 100% there's no credit at all. Once the IRAs open they're at 425.5%, over the cliff. So they get $0 in credits from 2027 to 2036, against $359,550 in benchmark premiums. Plan Mirror works this out every year, for every plan.

Testing ChatGPT's fixes

ChatGPT didn't miss the bridge. By its third reply it had a heading for the couple's “early-retirement funding gap”, and across that reply and the next two it offered three ways to close it: a 72(t) schedule of fixed IRA payments, Roth conversions, and spending a little less, “such as to $85,000 annually for a few years”. What it couldn't do was say which one works.

Plan Mirror spots the same hole, and goes further. Run this plan and its assistant calls 2031 an access gap, money that's there but locked, and suggests fixes, starting with a 72(t) schedule. A 72(t) schedule, Roth conversions and your spending level are all plan settings you can ask for in the conversation, so we ran ChatGPT's three fixes through the same five tests:

• A small 72(t) schedule. $25,000 a year from an IRA, penalty-free, from 2027 through 2031. Our stand-in for ChatGPT's “small, carefully structured 72(t) arrangement”; the real amount and end date come from the IRS rules.
• Roth conversions. $40,000 a year in today's dollars, growing with inflation, moved from the IRAs to Roth, from 2027 through 2031.
• Spending less. $85,000 a year all-in instead of $90,000. We cut it for the whole plan, not just a few years.

The 72(t) schedule does the most. It opens a slice of the IRAs from the first year, so less has to come from the brokerage account and the Roth, and the money they can reach lasts through the bridge. On steady returns there's no shortfall at all, and it's the only fix that also survives both stress tests: $1.7 million left with spending 10% higher, $512,000 with returns 2 points lower. On history, 90.3% of start dates make it, against 64.1%; on random futures, 83.5% against 47.7%. The cost on steady returns is small: they finish with $2.5 million in today's dollars instead of $2.6 million. It also fixes the insurance floor. The payments lift their counted income to 177.4% of the poverty line in 2027, so they qualify for $143,619 in credits. The catch is rigidity: the payment comes out every year until 2031, in good markets and bad, and a change the rules don't allow generally brings back the 10% penalty on every payment so far.

Roth conversions don't help the bridge. Before 59½, converted money has its own five-year wait before it can come out penalty-free, so in these runs the first conversion opens in 2032, the same year as the IRAs. Meanwhile the tax on each conversion comes out of the money they can reach. They still run short in 2031, and history gives 61.6%. ChatGPT itself cautioned against large early conversions. They do one useful thing: the extra income brings $322,336 in insurance credits. These runs hold spending fixed, so the credits don't show up in the success rates.

Spending less helps, but less. At $85,000 the steady-returns run has no shortfall, and history gives 72.6%, random futures 59.4%. That's a lifetime cut, and it still does less than the 72(t) schedule.

ChatGPT had the right list. It just couldn't rank it, and the fix it offered as a contingency turned out to be the one that matters. In Plan Mirror, you can test each one and see.

Verdict

Can this couple retire in 2027? The total is big enough: on steady returns they finish with $2.6 million in today's dollars, and every failing history fails before Social Security. Their access to the money is what breaks the plan. That's a specific, fixable problem: a small 72(t) schedule takes history from 64.1% to 90.3%. ChatGPT named that fix and called it a backup; the tests make it the first thing to set up, with a tax professional working out the real amount. It's exactly the kind of answer a one-off chatbot model gets close to without ever pinning down.

ChatGPT is a good thinking partner. It asked sensible questions and knew the rules. But a retirement plan is a 40-year bet with your savings, and it deserves an engine that gives the same answer every time, runs the tests you'd never think to code, and shows you the year it breaks. That's what Plan Mirror is for.

Plan Mirror uses AI too: it turns your conversation into plan inputs and explains the results. The numbers come from the engine, never from the AI.

The fine print

• The 59½ date. We didn't give Plan Mirror the couple's birth month, so it opens the IRAs at the start of 2032, the year they turn 60. Born in 1972, they reach 59½ somewhere between July 2031 and June 2032, depending on the month. Given a birth month, Plan Mirror opens the IRAs in the quarter they reach 59½. If theirs falls in 2031, these runs give them less access than the law does, and 2031 looks worse here than it would for them.
• Insurance. Premiums are national-average benchmarks, not county prices, and Plan Mirror doesn't model Medicaid below 100% of the poverty line. Spending is all-in, so premiums come out of the $90,000 and don't change the runs.
• The 60/40 test. We ran 60/40 as one mix across all the couple's accounts. ChatGPT's fuller version held the bridge money in cash and Treasuries and the stocks in the IRAs; we didn't test that version.
• One conversation. ChatGPT's answers vary between conversations and models. This page quotes one conversation, held on November 6, 2026, in a Temporary Chat in ChatGPT's web app.

Every figure on this page comes from 5 recorded cases run together in Plan Mirror under 2026 tax rules. Each was run through the plan with its two stress tests, a historical replay of U.S. stock returns and inflation from 1926 Q3 to 2026 Q2, and 1,000 Monte Carlo paths. It is an illustration, not financial, tax, investment or legal advice. See the methodology.

Related scenarios

• 72(t) vs the rule of 55: one couple at 55, three ways into the 401(k)
• Can I retire at 55 with $2 million?
• The ACA subsidy cliff at 58: one IRA-heavy couple, three plans