I Tested Claude Fable 5.1 - One Task Froze It
We ran Claude Fable 5.1 through language tests on launch day. The new model hung on a simple grammar task and never returned an answer.
Update (9 Nov 2026): Claude Fable 5.2 and Mythos 5.2 are out - prices, access and benchmarks are in my Claude Fable 5.2 guide.
👁 120 reads
We ran Claude Fable 5.1 through a set of language tests on launch day. The new model hung on a simple grammar task and never returned an answer - the same problem its predecessor had.
Every model launch follows the same ritual: the company shows charts where the new version beats the old one, and independent measurements appear a few days later. In between is a gap where nothing is certain. We decided to fill it our own way and ran Claude Fable 5.1 through a set of language tasks on launch evening - the same tests we’ve been using since mid-August.
The method is boring, and that’s exactly why it works. Six tasks, each with a clear-cut answer: declension of foreign proper names, collective numerals, capitalization, participial phrases, commas, and rewriting bureaucratic language into plain text. Same temperature of 0.2, same response limit, same time limit. Alongside Fable 5.1, we tested its predecessor Fable 5 and Opus 5, the model the same company sells for half the price. We ran these writing tasks in Polish, a heavily inflected language, making this a useful test of non-English language quality.
The Task the New Model Couldn’t Finish
The most interesting thing happened on task two: numerals. Fable 5.1 didn’t respond. It didn’t make a mistake, didn’t refuse - it simply returned nothing for three minutes until we closed the connection. Zero output tokens, zero cost, no response.
This isn’t a one-evening fluke. When we ran the same set in August, Fable 5 behaved exactly the same way: it hung on the proper name declension task and stayed frozen until the five-minute limit expired. In other words, this model family has a repeatable problem with freezing on short language tasks, and the new version hasn’t fixed it. In the same run, both Fable 5 and Opus 5 answered the numerals task without hesitation in under ten seconds, and both answers were correct.
Worth noting: we’re talking about a task that elementary school students do. Five children, two boys, four women, three students passed the exam, seven cars. There’s no trick here beyond the language itself.
Where It Did Answer, It Performed Very Well
The rest of the test looks favorable for the new version. Fable 5.1 solved the remaining five tasks correctly and was the fastest of all three: averaging 6.5 seconds per response, compared to 7.2 seconds for Fable 5 and 8.2 seconds for Opus 5.
The clearest quality difference appeared in the capitalization task. Both Fable models corrected the sentence according to standard rules: mayor of Warsaw lowercase, city office as an unofficial name lowercase, Mr. Director lowercase, and Ministry of Digitization and Christmas uppercase. Opus 5 left City Office and Mr. Director capitalized - exactly the mistake most common in official documents.
A detail that amused us: in the comma task, the input text lacked diacritical marks, because that’s how we’ve always provided it. Fable 5.1 was the only one that restored diacritical marks throughout the entire paragraph on its own, though no one asked it to. Fable 5 and Opus 5 politely left the text as they received it. Overzealousness, but in a good direction.
What It Cost
The entire six-task set cost us 7 cents on Fable 5.1, almost 9.6 cents on Fable 5, and just under 6 cents on Opus 5. Fable 5.1 came out cheaper than its predecessor despite identical token rates, because it simply talked less: when correcting participial phrases, it closed at 88 tokens where Fable 5 needed 257 and Opus 5 needed 358.
There’s an exception that spoils this picture, however. In the capitalization task, Fable 5.1 used 675 output tokens to return an answer character-for-character identical to what Fable 5 gave in 359 tokens. For exactly the same result, we paid almost twice as much. For one sentence it’s pennies, but for thousands of queries per day it’s a budget line item.
What This Means for Model Selection
Fable 5.1 costs $10 per million input tokens and $50 per million output tokens. Opus 5 from the same stable costs $5 and $25 - exactly half. In our test, the more expensive model won one task, lost one through lack of response, and was faster by a second and a half per query.
This isn’t a result that justifies a twofold price difference for writing and editing in non-English languages. If you do this daily, the cheaper Opus 5 will do the work just as well, and with longer texts will leave half the bill in your pocket. Fable 5.1 defends itself elsewhere - in code and long research tasks, which this set deliberately doesn’t touch.
One caveat we set for ourselves: this is a launch-day measurement, on six tasks and one run each. This is not a benchmark with hundreds of repetitions and doesn’t pretend to be. It shows how the models behaved on our desk the first evening, and it shows one thing that no official materials will show: that the new version can freeze on a task for an eighth-grader.
We’ll repeat the same set in a few days when the launch traffic dies down and check whether the freezing on numerals disappears. If it stays, it will mean the problem sits in the model, not in server load.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →Frequently Asked Questions
Does Fable 5.1 handle non-English text better than the cheaper Opus 5?
In the six-task test, Fable 5.1 won one task and lost one through lack of response, while being faster by a second and a half per query. We found the result doesn’t justify the twofold price difference for writing and editing work in non-English languages.
Why was Fable 5.1 cheaper to use than its predecessor when token rates are identical?
Fable 5.1 generated shorter responses - for example, when correcting participial phrases it closed at 88 tokens, while Fable 5 needed 257 and Opus 5 needed 358. The exception was the capitalization task, where Fable 5.1 used 675 output tokens for an answer identical to what Fable 5 gave in 359 tokens.
What was the problem with Fable models freezing?
Fable 5.1 returned no response to the numerals task - it didn’t make a mistake or refuse, it simply generated no output tokens for three minutes. An identical problem occurred earlier with Fable 5, which hung on the proper name declension task and stayed frozen for five minutes.
More on this from QUANTSLANT
guides·Midjourney Style Creator: Train AI for Your Visual StyleStyle Creator lets you build repeatable visual styles in Midjourney by clicking images, not writing prompts. Here's how to create and use custom style codes.