Anthropic released Claude Fable 5.2 on November 9. It is their most capable public model, and it costs twice as much per token as Opus 5.5.
I ran my monorepo bench on it. Fable tied Opus. I built two harder tests. Fable tied one and lost one. I had the post written: no test it wins.
Then I read the runs it lost.
The Sticker
| Per million tokens | Input | Cached input | Output |
|---|---|---|---|
| Claude Fable 5.2 | $8 | $0.20 | $40 |
| Claude Opus 5.5 | $4 | $0.20 | $20 |
Anthropic does not pitch Fable as the everyday model. Opus 5.5 stays the default in Claude Code on every plan.
— Anthropic, Fable 5.2 announcementFor most coding and knowledge tasks, we recommend that users start with Claude Opus 5.5, our strongest daily driver
Three Tests, One Wrong Answer
Same setup as before: real merged changes from a production TypeScript monorepo of about 3,700 files, replayed from a clean snapshot and graded by the tests that shipped with them. Claude Code 2.1.321, medium effort, the default for both.
| 52 runs each: passed every test | Fable 5.2 | Opus 5.5 |
|---|---|---|
| Six merged changes (20 runs) | 13 | 13 |
| Four large, fully specified changes (4 runs) | 4 | 4 |
| Seven bug reports, symptom only (28 runs) | 23 | 27 |
| Total cost at list price | $127.93 | $78.44 |
By the tests, Fable is no better and costs 1.6x. The tests were the problem.
- Test one has a ceiling. Nobody can score above 13.
- Test two is too easy. Both passed all 194 tests.
- Test three marked Fable down for being right.
Then I Read the Failures
Fable failed five bug-report runs. All five were correct fixes. An exact-match test rejected an extra field.
Two of those bugs sit in the same routing code. There, Fable made a change the ticket never asked for, in 6 of 8 runs. Opus never did. I called it scope creep.
It was the right fix.
- The human fix shipped without it. It dropped an option after an outside system replied, so the fallback the ticket said to keep never showed up. The ticket was bounced.
- Four days later, a second commit added it. Fable made that change on the first try.
- My tests were written for the first fix. They can’t see the hole. They can see an extra field.
Then I had both models review finished changes. Both caught every known defect. Only Fable caught this one.
Opus wrote the fix the tests wanted. Fable wrote the fix the ticket wanted. Reviewing Opus’s passing fix, Fable flagged the hole in 4 of 4 runs. Opus flagged it in 1 of 4.
The Fallback Hurts the Innocent
Fable has classifiers that block risky security and biology work. A blocked request is not refused. It is answered by an older model: Opus 4.8 for security, Opus 5 for biology. In Claude Code the session then stays on that model until you run /model.
Now think about prompt injection. Your agent reads a poisoned web page. You did nothing wrong. The classifier sees attack text and hands the request to Opus 4.8.
Anthropic tested this in the system card, with an attacker making 200 attempts per scenario against a coding agent:
| Shipped safeguards on | Attack success | Fell back |
|---|---|---|
| Fable 5.2 | 28.16% | 99.5% |
| Opus 5.5 | 11.13% | 64% |
| Fable 5.1 | 8.70% | 60% |
Fable answered 43 of those attacks itself. It lost none. Almost everything else went to Opus 4.8, which lost 28.56% of the time.
— Anthropic, Fable 5.2 system cardFable 5.2’s higher attack success rate therefore reflects a higher fallback rate (99.5%, compared with 64% for Opus 5.5 and 60% for Fable 5.1) rather than a difference in the models’ own robustness.
The safest model Anthropic has built is swapped out at the moment an attack arrives. I saw the same pattern in Fable 5.1. It is bigger now. To stop the silent switch, open /config in Claude Code and set “Switch models when a message is flagged” to ask each time.
What It Cost My Plan
I started the day with 5% of my weekly Fable allowance used, on the $200 plan. After 86 Fable runs ($169 at list price) and the research agents for this post, it was at 41%.
X is full of worse numbers: whole five-hour windows gone in one prompt. Many of those name the $20 plan, a higher effort level, or Fable subagents that also run Fable.
What This Doesn’t Show
- One bug is one bug. The hole Fable caught is a single case in a single codebase. It is the only place the models differed in substance.
- Fable’s fix was not the team’s fix. It made one of the two changes in the second commit, in a narrower form.
- Medium effort only. Anthropic’s gains are at high effort and above, on long tasks. Most of my runs took 1 to 14 minutes.
- Review costs more too. Fable’s 34 reviews cost $41.01. Opus’s cost $17.06.
- The injection test is an attack. Fallback on normal work is rarer, but not always rare. Vals counted it on 1.88% of its index, and on 20% of its code migration benchmark.
What I’d Do
- Opus 5.5 for routine tickets. Same test results here, 60% of the cost.
- Fable where a wrong fix is expensive. Money, routing, anything that talks to another system.
- Fable as the reviewer. It found everything Opus found, plus the one that mattered.
- Turn off auto-fallback. In
/config, set “Switch models when a message is flagged” to ask each time. A session that stops is safer than one that quietly moves to an older model. - Don’t trust a tie. My tests said these models were equal. The tests were the limit.
Related Posts
