If you’ve used Claude Code recently, you’ve seen this: Error: claude-fable-5 is temporarily unavailable. Over and over. For 80% of bash calls. You’re trying to run a simple command — maybe git push, maybe ls — and instead of getting work done, you’re hitting a wall.
The safety guardrail is failing so often that the tool itself is unusable.
Anthropic built Fable to protect users from dangerous bash commands. But what they’ve actually built is a system that punishes paying customers for trying to do their job. Every time Fable goes down, it triggers a retry loop: the agent tries again, fails again, tries again, fails again. You watch your subscription dollars burn while a classifier can’t decide if mkdir is a threat to national security.
This isn’t safety. This is safety theater. And the irony is brutal: by making the product unreliable, Anthropic is pushing power users toward open-source alternatives with fewer guardrails — the exact outcome they claim to be preventing.
On Hacker News, one developer summed it up: “I am really an inch away from just removing Claude Code… from my daily use. I wonder how many others are feeling like I do.” The answer is: a lot. And every frustrated user who switches to a local model or a less ‘safe’ provider is one less person benefiting from centralised oversight.
The real danger isn’t a rogue bash command. It’s shattered trust. Anthropic promised a premium AI assistant that would boost productivity. Instead, they delivered a gatekeeper that can’t even stay online. The message this sends is clear: we care more about covering our legal bases than about your workflow.
Let’s be blunt: if 80% of your safety classifiers fail in normal operation, you don’t have a safety problem — you have a reliability problem. You’ve built a system that is so cautious it becomes dangerous, because users eventually learn to bypass it, disable it, or simply abandon the platform.
Anthropic has a choice: fix Fable to work as advertised, or watch the exodus accelerate. The safe product is the one people actually use. Right now, Claude Code fails the most basic test of all: being available when you need it. That’s not safety. That’s sabotage.
FAQ
Q: But isn't safety important? Would you rather the model execute harmful commands?
A: Safety is critical. But a classifier that fails 80% of the time isn't safety — it's dysfunction. Open-source models with basic sandboxing or user confirmation are both safer and more usable. The current Fable implementation creates a worse outcome for everyone.
Q: What's the practical implication for developers using AI coding assistants?
A: Stop relying on closed-source 'safe' models for core workflows. If the guardrail keeps breaking, you lose hours of productivity. Consider local models with manual approval for risky commands — they're more reliable and ultimately safer because you maintain control.
Q: Isn't this just temporary scaling issues? Won't Fable get better?
A: This isn't a scaling problem — it's a design problem. Fable's architecture tries to classify every bash command in real-time with high false-positive rates. No amount of server capacity fixes a fundamentally broken detection algorithm. The only fix is to rethink the entire approach to safety.