On 10 November, Claude Code on my Max plan switched its default model to Fable 5.2. Fable costs twice what Opus 5.5 does per token, and the model picker calls it the one “for your toughest challenges”. The next day the default was Opus again, but by then I had raced the two on an everyday job and on a real bug that Cognition says Opus 5.5 fails at every effort level. Fable failed that bug 11 times out of 11.
📺 Watch the video: Fable 5.2 vs Opus 5.5: I Paid Double and Got the Same Bug
Two models and one price list
| Per million tokens | Input | Output |
|---|---|---|
| Opus 5.5 | $4 | $20 |
| Fable 5.2 | $8 | $40 |
Both ran at medium effort, which is where Claude Code starts and where most people leave it. I used Claude Code 2.1.322 in -p mode on my Max 5x plan and my Mac mini M4, one run at a time, with the two models taking turns. The plan doesn’t bill per token, so every dollar here is Claude Code’s own count at API list prices.
A hidden test the models never saw graded every run. To compare them I use cost per passing run: everything a model spent, divided by the runs that passed every check.
Test 1: an everyday job
The task was the voucher step of the shop till from my effort staircase post, 5 runs each, 36 hidden checks.
| Passed | Per passing run | Median time | |
|---|---|---|---|
| Opus 5.5 | 5 of 5 | $1.60 | 7:57 |
| Fable 5.2 | 5 of 5 | $2.95 | 6:42 |
Fable got the same result a little quicker, for almost twice the price. On my plan the gap was wider. Each Fable run moved my session bar by 7 to 10 percent and each Opus run by 1 to 4, about three times as much.
Both also handled the big orders the task never mentions, 10 runs out of 10. In my last two videos Opus at medium failed those 8 times in 9 through the SDK, so something changed, and I don’t know what yet.
Test 2: the bug Opus fails
On 27 October Cognition, the company behind Devin, published a set of hard coding tasks called FrontierCode 2.0. One of them is a real bug in Mock Service Worker, the library that fakes network requests in tests. In Cognition’s runs Opus 5.5 passed 0 of 5 at medium, high, extra high and max.
Mock Service Worker lets you answer a request with a generator, a function that hands back one value at a time: pending, pending, then done. The bug report says that after resetHandlers() the generator doesn’t start over, so the next test gets “done” straight away. That report plus “Fix it.” was the whole prompt.
The report never mentions that a generator can have a finally block, which is where cleanup goes. If the library drops the old generator without closing it, that cleanup never runs.
Cognition doesn’t publish its grader, so I wrote my own hidden test: 3 checks that the generator starts again after resetHandlers(), 4 that the old generator gets closed so its finally block runs, and the repo’s own 100 tests. The library’s real fix (PR 2725) passes all of it.
| At medium | Passed | Spent in all |
|---|---|---|
| Opus 5.5 | 1 of 5 | $2.19 |
| Fable 5.2 | 0 of 5 | $4.39 |
All nine fails, from both models, wrote the same two lines that drop the old generator without closing it, and every one of them had its own tests green. The one pass was an Opus run that added return() to close the old generator first.
Test 3: more effort
The picker says Fable is for the tough problems, so I turned its effort up.
At high, Fable went 0 of 5 for $8.25. It thought about four times as much as at medium and wrote up to 36 tests of its own.
At max I ran it once. It took 34 minutes, wrote 46 tests, started a second agent to review its fix, and still failed, for $17.71. That one run took my session bar from 33% to 95%, so I stopped there.
In all, Fable went 0 for 11 on the bug and spent $30.35 at list price. Opus spent $2.19 on its five runs and fixed it once.
What I took from it
This is one everyday job and one bug on my machine, with small samples. On these two Fable wasn’t worth double. It matched Opus on the everyday job for nearly twice the cost, and on the hard bug it made the mistake Cognition saw Opus make, at every effort level I tried.
I still haven’t found the task where Fable earns its price. If you’ve found a bug Fable fixes and Opus doesn’t, send it to me in the video’s comments and I’ll race them on it.
Your default may change again without you asking, so run /model in Claude Code now and then to see which model you’re on.
Race them yourself with which-claude
which-claude 0.5.0 runs this comparison in one command with your own check and shows the cost per passing run. It has Fable 5.2’s prices and a --task-file option, and example/msw holds this exact bug with my hidden test (steps in the README):
which-claude --models opus,fable --check "npm test"
Code & video
- 📺 Full video: Fable 5.2 vs Opus 5.5: I Paid Double and Got the Same Bug
- ⚡ The Short: youtube.com/shorts/CDzBxEEx4Tw
- 💻 which-claude: github.com/thecoderlog/which-claude
- 🐛 The bug: Cognition, FrontierCode 2.0 and Mock Service Worker’s fix
- 📚 Docs: Pricing, Claude Code model configuration
- 🔁 Last video: I Gave Cheap Claude an Opus Advisor. Did It Beat Opus?
If this was useful, subscribe to The Coderlog on YouTube for the next test.