Fable 5.1
Anthropic released Fable 5.1 this week. My initial reaction to the release was this comment I made in the Third South Slack:
More seriously: I'm excited to see the reduced cost, but have a very hazy time of describing the shape of work that Fable 5 was unable to complete, which is not to say that it didn't exist, but that it's hard to even conceptualize a step function improvement because the line between "LLM can do this" and "LLM cannot" feels so amorphous. I suspect their lack of concrete examples here is proof that this is a broader issue.
Later that afternoon, I fired up 5.1 in much the same way I had fired up Fable 5 for the past few weeks, both on Conductor and, an hour or so later, on the Claude Code webface.
When I fired up the webface, it prompted me -- in a full-screen modal dialog -- to try out Fable 5.1, even though Anthropic had all the information it needed to know I had already tried it out.
Then, later that evening, I wanted to look at one of the long-running tasks I had kicked off before leaving work. So I opened up Claude on my phone, at which point it announced -- in a full-screen modal dialog -- the release of Fable 5.1.
The two options to dismiss this dialog were "try it now" and "maybe later."
This little papercut is not novel in any sense: it's happened for every model release from Anthropic in recent memory, and is a common genre of problem across all sorts of applications.
Anyway: Fable 5.1 seems good. It is faster and cheaper for the work I've used it for than Fable 5. And I am quite confident that someone could probably feed it the contents of this blog post and it would fix the issue I am whining about. But every time I hear a frontier lab employee mention that software engineering is a solved problem I pretend they say that cooking is a solved problem.