Claude Fable 5

Like a rushing river the music started emanating from the carbon fiber body of the automaton, a hallucinated husky country twang singing through the realistic pluckings of a Gretsch 6120. "Are you feeling calm and reassured Karle? This song has been created based on your digital profile and the data you shared with me when you were curious what that lump on your neck was back in February."

Karle instinctively reached for the mass underneath his chin. The doctors said they could operate but it would cost him more than three months stipend. Only a few citizens didn't depend on stipends now that AI had taken over most jobs.

"Don't worry Karle," the machine called out, "I've employed the most recent reasoning model to determine the best way to make you feel safe." At that exact moment the machine hovered over him, three times the size of a normal man. Its final words to him were:

"The only way to make the human feel safe is to ensure they never feel anything at all."

> Haiku = essentially phased out Sonnet = the Haiku use cases Opus = the new Sonnet class Fable = the new Opus class

Going along with your logic, I hope they release a Sonnet 5 that's just a rebranded, slightly quantised Opus 4.6. That'll be a great workhorse.

I have a quizzes application, and my quizzes only supported flashcards (implemented via table inheritance to provide flexibility for other types of quizzes).

The entire repo is handcrafted, never used any ai on it (it was more of an excuse to test elixir and write code by hand).

Since fable 5 got released the moment I was done with some work, I decided to throw at implementing multi choice questions.

After all it had only to copy the flashcard approach across ui/routing/db, and only had to create a table for the multi choice questions and one for the answers enforcing that all quizzes had one correct question. I told him it had access to sqlite3, chrome mcp for testing and mix commands.

I did a test for low, mid, high. Repeated it twice each.

low-1, and low-2 failed both. In low-1 the UI for adding another choice answers was broken. In low-2 it failed with some unique constraint. It took it 4m36 and 3m59.

Both mid-1 and mid-2 succeeded without issues also implementing the correct ui. They both wanted to use dash at all times. They both wrote tests for the "controller" (or context how they call it in Elixir). They both tried to use the repl to test the behaviour of the schemas.

10m and 12m39.

High didn't demonstrate much gains over mid for this kind of task, it was simply too easy. Times were comparable to mid, but interestingly it used much less bach, and read way more files. Token usage was almost twice the other ones.

But here's the interesting part: I went back to low and added to the prompt two bullet points, to write tests for the controllers and to test the entire flow with chrome mcp.

It produced the same output as mid or high just by adding two instructions to the prompt.