This is maybe an unpopular take, but I don't think this matters for Anthropic. I don't think their immediate goal is to get users onto their biggest and best mo

This is maybe an unpopular take, but I don't think this matters for Anthropic. I don't think their immediate goal is to get users onto their biggest and best models. Sonnet is more than enough for many average users and their use cases. At this point, Fable is really just an experimental model (as it should be). It can do very useful things, but is it a broadly good general purpose model? Definitely no. Most users don't have problems hard enough for Fable outside of coding very large and complicated projects. Most users don't have 45 minutes to accomplish a task Sonnet can do well enough in five. There's not a PowerPoint in the world where Fable is the right tool to build it. The secret sauce is going to be in letting Sonnet decide to delegate to Opus and Fable when they're the right tools for the job. But you can't train a model to do that until the bigger/better models exist and you can observe how your users actually take advantage of them. I'd bet money that's exactly what the rlhf going on at Anthropic looks like right now. Economically, it makes sense. Sure, on paper you want users burning as many tokens as you can. But pushing users into burning tokens and taking a long time and getting a meh result is far worse then giving them the "fast and good enough" solution that occasionally burns more tokens automatically when the problem requires it, and getting a higher quality result out when you do. From an infrastructure capacity perspective, this is the dream: you stop measuring cost [for Anthropic] per token and measure cost per outcome, allowing you to use less hardware to accomplish the same abstract units of work.