Anthropic's Powerful New AI Arrives With Hidden Filters — and a Public Apology

Anthropic launched Claude Fable 5 this week, billing it as the most powerful AI model it has ever made widely available. According to The Decoder, it is the first model in a new tier Anthropic calls "Mythos-class" and leads nearly every major benchmark, including a 95 percent score on SWE-bench Verified — a test of AI coding ability. The same outlet reports it costs twice as much as Anthropic's previous top model, Opus 4.8, at $10 or $50 per million tokens depending on usage.

But the launch quickly ran into controversy. According to The Decoder, strict safety filters block roughly nine percent of all requests. The Verge reported that the model refused to answer basic biology questions — the kind, the outlet noted, a high schooler would be expected to handle — despite Anthropic marketing the model's science capabilities. TechCrunch reported that cybersecurity researchers are openly complaining the guardrails are too strict to be useful for legitimate security work.

According to MarkTechPost, Fable 5 and the more restricted companion model, Mythos 5, share the same underlying architecture but differ in what their safety classifiers allow. Mythos 5, which remains limited in availability, has cybersecurity restrictions lifted through a program called Project Glasswing.

Anthropics's response drew its own criticism. According to Yahoo Tech, the company apologized for not disclosing the content restrictions upfront — but the fix reportedly comes with caveats.

More trouble arrived from an unexpected direction. According to The Verge, Microsoft internally restricted employee use of Fable 5 over concerns about Anthropic's data retention requirements. Separately, Microsoft AI CEO Mustafa Suleyman told The Verge's Decoder podcast that Anthropic's suggestions about Claude's potential consciousness, written into the model's guiding instructions, are "really, really dangerous."

The episode matters because it shows that as AI models grow more capable, the tradeoffs between safety and usefulness are becoming harder to hide — and harder to get right.