Is Fable just overpriced hype?

I've replaced Fable with Opus in my main bot. I'd chosen Fable because, in theory, it was the best. If I was going to work with several bots powered by different models, it seemed reasonable to reserve the supposedly most capable model for that main bot. The logic was impeccable. The experience, rather less so.

I work with Opus, Sol and Fable in parallel. And my experience with the last of these has been poor: it seems very capable, answers with enormous confidence and, when I look at the result, I find mistakes, delays and things it has forgotten. In my work, I've seen it perform on a par with Sol 5.1 or worse. That's what led me to switch.

The question in the title expresses that disappointment. If I buy into the promise of something extraordinary and end up wrestling with ordinary problems, I wonder what I'm paying for. The fare may be first class; I'm still waiting for my suitcase.

Confidence doesn't get the work done

What bothers me most about Fable is the gap between the impression it makes and what I can actually use in the end. When someone seems to understand everything, you expect that understanding to show up later. For the assignment to move forward. For the important things to stay in view. For the result to match what you asked for.

A confident tone encourages you to lower your guard. I want to do exactly the opposite: look at what it delivers. The confidence with which a model speaks is part of the conversation; quality lies in what I can use when it's finished. Confusing the two is costly, even if the answer is beautifully written.

I don't need an AI that wins an argument with me about how capable it is. I need it to do the job well. And if it needs correcting, that correction should show up in the result. The user's patience shouldn't be listed among the system's unlimited resources either.

In a work tool, forgetfulness and delays matter. A flash of brilliance can be impressive, but trust is earned through consistency. I'd rather have help that delivers reliably than help that leaves me wondering how much of its brilliance will survive to the finished work.

The price has to be justified

There is a specific difference in their rates. Anthropic, the maker of both models, charges standard API rates of $10 per million input tokens and $50 per million output tokens for Fable 5.1; for Opus 5.5, the rates are $4 and $20. These are usage rates, not the final cost of an assignment or a subscription fee. Official Anthropic pricing.

Paying more can make perfect sense. If a tool solves something the others can't, or helps me finish an important piece of work to a higher standard, the difference may be money well spent. What I find hard to accept is paying for a superiority I can't see in my own work.

The time spent reviewing counts too. If an answer means revisiting the assignment, reminding it of what was missing or waiting for another round, the cost doesn't end with the tokens. My attention has value. The model that looks cheaper can turn out to be expensive; the most expensive one has even more to prove.

Fable may be excellent at other tasks, including extended reasoning, but in mine it hasn't justified the role I gave it.

That's reason enough to change my choice of tool. There's no need to hold a referendum on artificial intelligence every time a tool stops being worth it. Nor am I obliged to keep using it to prove I'm capable of appreciating its greatness.

Even its maker recommends starting elsewhere

Fable 5.1's own documentation recommends starting with Opus 5.5 for most workloads and turning to Fable for demanding needs or when your own evaluations with Opus fall short of what's required. That strikes me as much more useful advice than chasing a crown. Official Fable 5.1 guide.

The best model in a table may not be the best model for your work. The table reflects the tests behind it; your working day asks different questions. Was the assignment finished? Were the instructions followed? How much is left to fix? These questions don't look particularly dazzling in a presentation, but they matter quite a lot when you want to close your laptop.

I've already written about my experience with Claude Opus 5.5 and GPT-6.1 Sol in real projects. What interests me here is the extra expectation I placed on Fable: believing that choosing the supposed pinnacle of capability would spare me the problems I already knew. My experience has been different.

A ranking can guide me in choosing what to try. A model has to earn its place in my work afterwards. If I keep a model for its reputation while its output disappoints me, I'm defending my purchase instead of evaluating my tool.

Make it earn its place

My advice is to test it on tasks that actually matter in your working day. A familiar assignment, with a result you know how to judge. Then review what's been delivered, counting the corrections and the waiting time too. The first answer may be seductive; what you need to know is whether you can make use of the work.

It helps to decide beforehand what you'll consider good enough. If you change the standard every time your favourite stumbles, you'll always find a reason to keep it. And if a simpler model does the job, there's no prize for continuing to pay for the most expensive one. Or for becoming its defence lawyer.

Is Fable just overpriced hype? For me, the costly part has been expecting a difference I haven't found in my own work. I'm not accusing the product or its maker of intending to deceive. I blame myself for giving too much credit to the promise of capability before demanding results.

That's why I've removed it from my main bot. If it earns that place back by doing useful work, it can return. In the meantime, I'd rather my AI sounded less assured and left me with fewer things to fix.