Anthropic's Claude Fable is back — and users say it's answering "no" to almost everything
After regulators lifted the ban, the returning model keeps handing tasks off to a weaker sibling. Anthropic says its safety net is just set very wide.
Key points
- The US Department of Commerce lifted its ban on Anthropic's Claude Fable model, and the company relaunched it to all paying users this week.
- Users on the $100-a-month Max plan can only spend 50% of their weekly usage on Fable, and from 7 July it moves to pay-as-you-go credits only.
- Developers say the restored Fable frequently refuses ordinary coding tasks and silently switches them to the older Opus 4.8 model.
- Anthropic says the model itself has not been weakened, but that a large "safety margin" in its guardrails is catching far more prompts than before.
Anthropic's most powerful model is back. Its users are not thrilled.
Claude Fable — the flagship artificial intelligence assistant that US regulators had ordered off the market — returned this week after the Department of Commerce lifted its ban. On paper, that is a win for anyone paying Anthropic for access. In practice, the reception has been rough.
The first surprise was pricing. Fable is included in the Pro, Team and $100-a-month Max subscriptions, but each account can only spend half of its weekly usage allowance on the model. After 7 July, even that ends. Fable will move to a pure pay-as-you-go system, where users top up a credit wallet and are billed per request.
The bigger complaint is what happens when you actually try to use it.
Why are users saying the model feels "nerfed"?
Because it keeps refusing to do the work and quietly hands the job to a weaker model instead.
On Reddit and in developer forums, paying customers say the relaunched Fable behaves like a jumpier, more cautious version of its old self. It bails out of tasks it used to handle, and it does so visibly: Claude tells the user it is switching, then loads Opus 4.8, an older Anthropic model, to finish the job.
"The new guardrails are kicking in on way too many tasks and falling back to Opus 4.8," one user wrote. "This is not the model that got banned."
Guardrails, in AI-speak, are the automatic filters that stop a model answering things the company considers risky — anything from weapons instructions to sensitive personal data. The problem, users say, is that Fable's filters now trip on things that are not risky at all.
One developer said Fable "didn't even let me search for dead code without switching to Opus." Dead code is simply old, unused programming that engineers routinely clean up.
Another said the model was effectively unusable for low-level programming work. Files written in C, C++ or Rust, anything touching the Win32 API — Microsoft's core Windows toolkit for developers — and anything involving computer memory seemed to set off the filter. So did files containing everyday words like "security," "vulnerable," "unsafe" or "hook."
Those are ordinary terms in software engineering. A developer auditing their own product for weaknesses uses them constantly.
Has Anthropic actually weakened the model?
The company says no, and BleepingComputer, which first tracked the user complaints, reports the same. What has changed, according to Anthropic, is the size of the "safety margin" around the model — how much benefit of the doubt the filters give a borderline prompt.
A wider margin means more false alarms. Ask about a memory bug in your own code and the system may decide it looks close enough to hacking to bounce you to Opus.
Threat Vectr tested the model and saw the same pattern: Fable rerouting requests that had no obvious safety angle at all.
Anthropic has not publicly acknowledged the false-positive reports. It is a reasonable bet the company is aware and tuning behind the scenes.
What should paying users do in the meantime?
Two practical things.
First, watch your usage. The 50% cap and the July shift to credits mean it is easy to burn through an allowance faster than you expect, especially if half your prompts get rerouted and you retry them.
Second, if your work touches security research, systems programming or anything involving the words that seem to trigger the filter, expect friction. For now, Opus 4.8 is doing more of the heavy lifting on Anthropic's platform than the marketing suggests.
