TBH Fable tends to over think problems quite often - spins a lot as well.😅

Only makes sense when your go-to practical intelligence model fails on a hard problem

GPT 5.6 is poised to make it's debut as your everyday model

I am testing Fable right now. On one occasion it helped me find really good bug in my project.

On another one it missed some architectural planning and GPT 5.5 XHIGH was better, here:

I am testing Fable on coding now. IT IS REALLY SLOW. Lets see if produced code is any better and if its worth using it in coding role

@bindureddy Indeed, its a model for planning things or reviewing repos, for everyday driver opus 4.8 xhigh is anthropic’s best option

Opus also feels better since update - i think they rolled out noticeable harness improvements alongside fable

@PawelJLisowski @bindureddy ??????

I would use sonnet 4.6 or GPT 5.5 to execute and drive

Opus 4.8 x high and Fable at Med are roughly equivalent

So why use Opus 4.8?

@bindureddy i think it overdoes the easy 90%, but never touches the hardest last 10%

@bindureddy curious what you're testing it on. I mean for ui/design tasks the overthinking weirdly helps, it fills in intent you didn't specify. but for quick scripts it's painful.

@jonascoleai @bindureddy tbh neither can junior devs. hand a vague spec to a junior and they'll build a framework. the overthinking isn't a bug, it's just what intelligence does when you don't constrain it.

@lokendra__sinh @bindureddy The UI/design point is underrated. Overthinking filling in unspecified intent is exactly what you want when the spec is vague. Same trait is a liability the second the task is well defined. The model can't tell which situation it's in.

@bindureddy yeah i guess it’s for super intelligent people to solve super hard unsolved problems

@bindureddy The overthinking is the feature, not the bug. Fable is built for the 10% of problems where you actually want it to spin. Using it as your everyday model is like bringing a structural engineer to hang a picture frame.

@bindureddy Yeah models just keep coming and going all the time. I guess AI has hit it's saturation level, mostly not that much improvement these days.

@bindureddy Fable just nailed something I've been working on for a few weeks in a morning. It's definitely a step up, how big I can't say.

@bindureddy Agreed. Over-reasoning models have high latency and spin wheels on simple tasks. For everyday deployment, edge-level execution with a solid practical model wins over pure compute depth every single time.

@bindureddy overthinking is a router miss, not a model trait. depth cap, early-stop on confidence delta, fall-through to cheaper when gain drops below threshold. is the abacus router doing all three or letting fable run hot?

@bindureddy Basically, it is REALLY good, but not ideal and definitely NOT "much better" than GPT 5.5 XHIGH.

5.6 will probably kill Claude Fable for everyday use

@bindureddy Interesting framing — "overthinking" vs "practical intelligence" is essentially the precision-recall tradeoff in LLM architecture. Curious how GPT 5.6's training objective differs to favor actionability over deliberation.

@bindureddy true, i dont think they ever recovered after 4.5/6 pre-nerfed opus.

@bindureddy gpt 5.6 really about to ruin the overthinkers 😩

@bindureddy I’m sorry you feel this way about the product, but I think both Claude 5 Fable and Chat GPT 5.6 are going to kick some ass! I think Chat GPT 6 and Mythos 2 are going to be light AGI and really make the beginnings of singularity spicy

@bindureddy fable5-medium is good