I'm looking for people who have given Fable a real shot and are unimpressed.
Tell me more about your experience with it below. Will follow up with a few specific questions
Theo Browne is reaching beyond promotional noise to collect detailed accounts from people who invested real effort in Fable yet came away dissatisfied, planning follow-up questions aimed at surfacing concrete shortcomings in the model released June 9 and redeployed July 1.
I'm looking for people who have given Fable a real shot and are unimpressed.
Tell me more about your experience with it below. Will follow up with a few specific questions
The outreach itself contains no listed complaints or benchmark shortfalls, so any patterns around long-horizon tasks, safety-classifier triggers, or harness integration stay unknown until replies arrive.
Because Theo maintains coding-focused tools and content, the collected feedback could clarify where Fable's safeguards or pricing at $10/$50 per million tokens create friction versus prior Claude releases.
Many users criticized Fable for being nerfed and unreliable after Anthropic changes that broke tasks and triggered refusals, while others praised its speed on complex work.
No Digg Deeper questions have been answered for this story yet.

1. Can you share some of your prompts? What was disappointing about the result? 2. What harnesses are you using? Claude code? Cursor? 3. What is your usual daily driver right now? 4. What's the biggest task you've ever seen an agent complete without human intervention?

@theo I’m neither impressed either.

@theo I use superpowers (https://github.com/obra/superpowers) and it's noticeable improvement over 4.8 is essentially non-existent. Essentially just burning tokens. Perhaps says just as much about superpowers elevating 4.8?

My issues with fable mainly lie with its ability to design proper user experiences. From my experience, it still lacks the ability to reason through human actions to properly optimize interfaces. I spend a significant portion of my time highlighting aspects of generated apps for correction. This holds true even with the use of skills, and across harnesses.

Having it create this fish tank took like 30 prompts.
The fish were teleporting when you tap, it decided they didn’t need to eat the pellets to refill hunger, it had giant text labels everywhere. It kept fighting me on decisions like “the fish need to go to the pellets to eat them”. It’s just been horrible.

@theo Is this unimpressive enough?
Great way to learn what Fable doesn't do.
I'm looking for people who have given Fable a real shot and are unimpressed.
Tell me more about your experience with it below. Will follow up with a few specific questions

@theo It gave me a warning about cybersecurity and i wasn't even coding i was asking about agentic marketing and persuasion

@theo looking for that too, specially with the prompts
it is really hard for me to imagine

@theo "it was a misalignment model and I really hope they deprecate it and then burn the weights." sounds familiar? 🙂

@theo The classifier thing is sensitive more than before, as expected.
I feel like we can only talk to Fable about design or the fucking weather at this point…

@theo Yeah its sus, i didnt hit any safeguards yet,it kinda feels like old fable but kinda feel sussy, not sure yet to buy another max plan or no,cuz it eats limits like crazy now.

@theo it's going to be people with memory enabled, tons of skills, a bloated redundant system prompt, a bunch of mcps, and other nonsense
all these people comparing new models to one another when their models are all lobotomized

@theo Well obviously it's not gonna perform as good when you add guard rails and fallback to opus 4.8 for certain coding tasks. That's America freedom country number #1 for you. Might as well start calling it China from now on.

@theo So far, so good. I basically get it to collate everything it needs thru other models, then it does the thinking. I just don't let it do the donkey work, even though it can go and get stuff for me. I like to make sure it gives a work to Codex and then works on everything after.

I haven't had much time yet but it's mostly just hitting safeguards just quicker just because there was mention on dll injections in the previous session i had.
Its how malware operates but for my use case which was in the same session wasn't for that but just prodding a running process on windows through ssh for some experiments im doing on a remote machine.

@ElliGrossman Nope, should be identical.

@theo I never asked it to do any coding. I asked it to do analysis and planning consistency more than anything when it comes to code, and honestly, it just felt like opus with more effort behind it and the compute to match, so it didn't feel slower.

I've been trying to perform cybersec work recently to secure some of my projects so it's really disheartening to have a model like Fable 5 just outright reject anything in the near vicinity of cybersecurity topics.
I am having better results for cheaper by using fusion models
-Over-refusals, blocks, and fallbacks on benign/borderline prompts -High cost and rapid quota burn -Perceived capability regression in practice for some

@theo i just got access, will let you know!