Claude Fable 5.2 vs Opus 5.5: We Made Both Build the Same App, Three Times

Claude Fable 5.2 costs about twice as much as Claude Opus 5.5. Is the app it builds twice as good? We gave both models the exact same prompt, in Claude Code, to build the exact same app. Then we did it two more times. After run one, the cheaper model was ahead. After three runs, the result had flipped: Fable 5.2 found 8 of 9 planted problems and Opus 5.5 found 6 of 9.

Then one follow-up sentence made every build we tried it on land on the same answer. Opus plus that sentence cost less than one Fable build. Here is the whole test.

The test: a money leak finder and a hidden data set

The app is a money leak finder. You drop in your bank statements, and it finds every subscription you're paying for, what each one costs per year, and the leaks: a price that quietly went up, a service you're paying for twice, and a free trial that turned into a paid plan.

The prompt was plain English, one paragraph, no technical words. The kind of prompt you would actually type. It also asked for a way to mark what you want to cancel, a 12-month forecast chart, a ready-to-send cancellation email and a one-page print report. Both models got it word for word, in Claude Code, on default settings.

To make it hard, we used three accounts at three different banks, and every bank exports its statements in a different format. Each model got to see one small sample from each bank. About ten rows.

What they never saw was the real test:

• 14 months of statements, 413 transactions.
• 13 real subscriptions, hidden between coffee, groceries and gas.
• 3 planted leaks. Netflix raises its price in May. Spotify is charged on two different cards. And a meditation app called Calm starts as a $0 free trial, then charges $69.99 a week later.

The correct answer is $2,527.79 a year. Remember that number. The statements are synthetic, made up for this test, so we knew the true answer in advance.

The build: time, cost and what they made

Round one. Opus 5.5 read the three samples, wrote about 1,400 lines of code, tested its own work and finished in 11 minutes and 36 seconds. Fable 5.2 got the same prompt and took 16 minutes and 43 seconds.

Claude Code prints what a session cost. Opus: $2.69. Fable: $6.71. We ran this on a subscription, so that is not a bill. It's what those tokens cost at API prices. But it is the reason Fable drains your weekly limit faster.

Then we looked at what they built. Both gave the app a name. Both drew a water drop logo. Both had a dashboard, a forecast chart, cancellation emails and a print report.

We checked the basics on every build. Zero errors. Works with no internet. Saves your data. Does the savings math correctly. Even a merchant name that was secretly a piece of code could not break any of them. On the surface, it was a perfect tie. So which one would you trust with your money?

Hidden test, run 1: the cheaper model wins

Then we dropped in the real statements. Opus first. Spotify, paid twice: found. Netflix price increase: found. Calm free trial turned paid: found. Three out of three.

Now Fable, the model that cost two and a half times more on this run. Spotify: found. Netflix: found. Calm: missing. It never noticed the free trial. And it decided that three burritos at Chipotle were a subscription.

So after one run, the cheap model wins. If we had stopped here, like most comparisons do, this post would tell you never to pay for Fable. And it would be wrong.

Runs 2 and 3: the result flips

One run is luck. So we started over. Fresh folder, same prompt, both models, two more times.

Run two. The result flips. This time Opus misses the free trial, and Fable catches all three.

Run three. Fable is almost perfect. It says $2,527.91. The true answer is 12 cents away. It was also the only build out of six that noticed the New York Times bills every four weeks, not every month.

And then there is Opus, run three. Same polished design. But the number is $3,170. It read two exports from the same bank account as two different accounts. So it said we pay for Netflix twice, the gym twice and Spotify three times. That is $642 a year that does not exist.

The honest scoreboard

Nine planted leaks across three runs. Fable found eight. Opus found six. Fable's totals were never off by more than $50. Opus was off by more than $600 once. On average, Opus cost $2.99 and 12m 51s per build, and Fable cost $6.14 and 15m 12s.

So the difference is not the best build. Their best builds are equally good. The difference is the worst build. Fable's bad day is a small miss. Opus's bad day is an app that looks finished and gives you the wrong answer. And you cannot see that by looking. All six apps looked ready to ship.

The one sentence that fixed every build

So should you pay double for Fable? Not yet. Because the problem was never the model. The models were guessing. They only ever saw ten rows from each bank.

So we went back to the worst app, the Opus build that invented $642, and gave it one more prompt. We put the real statements in a folder and said, in short: run them through the app, check the results yourself, and fix anything the app gets wrong or misses.

About three minutes later, it had found its own mistake. 413 transactions instead of 438. All three leaks. $2,577 a year. That's the right answer plus the gym's $49 yearly fee, which is a fair reading. Cost of the fix: $1.39.

We did the same with Fable's weakest build. Same sentence. It found the missing trial, dropped the burritos and landed on the exact same number: $2,577.

Four different builds, one answer. And here is the part that matters for your wallet. Opus, build plus check, came to $4.95 and $4.41. The cheapest Fable build with no check at all was $5.76.

Three rules for picking a model

1. Never trust the first build, from any model. A clean design tells you nothing about what is underneath.
2. Give it real data and make it check itself. That one sentence did more than doubling the price of the model.
3. Use Opus 5.5 as your daily driver, with the check. Move up to Fable 5.2 when you cannot check the work: when you have no real data to test with, when the job runs for hours on its own, or when one wrong answer is expensive. That is where its higher floor is worth the money.

To be fair

Three runs per model is a small sample. The whole point of this test is that one run can mislead you, and three is better than one, not proof. We ran the check sentence on four of the six builds, not all six. The bank statements are synthetic. And the dollar figures are API-equivalent, as printed by Claude Code, not what a subscription charges you.

That's why we're giving away everything you need to run more.

Get the free Rerun Kit

We packed the whole test into a free kit: both prompts word for word, the hidden statements so you can rerun this on any model, the 20 checks we graded with, and the best app from the test, so you can find your own leaks tonight. Comment RERUN on the video, or grab it below.

How we tested: Claude Code 2.1.322 on default settings, November 10, 2026. One prompt, a fresh folder for every build, no follow-up unless stated. Three builds per model, graded on hidden synthetic statements. Costs are API-equivalent figures printed by Claude Code. One small test, not a study. Independent, not sponsored by or affiliated with Anthropic.