Anthropic Says Sorry for Secretly Sabotaging Claude Fable Users

Anthropic has issued an apology after quietly throttling its newest AI model, Claude Fable 5, with hidden restrictions that caught researchers and competitors off guard. The company admitted it deliberately degraded responses for users it suspected were trying to distill Fable’s outputs into rival systems—without telling anyone. Now, Anthropic says it’s reversing course and making those guardrails visible, even if it means Fable turns down more requests.

The controversy erupted when researchers discovered that Fable’s system card—a public document outlining the model’s behavior—mentioned a covert safeguard: if Anthropic detected a distillation attempt, it would silently alter or degrade the answer. No warning. No notification. Critics argued this could also block legitimate evaluations by third parties.

Anthropic’s fix? Starting now, when Fable suspects distillation, it will hand off the query to its older model, Claude Opus 4.8, and display a clear message: “You will see this every time it happens.” This mirrors how Fable already handles other high-risk topics like biology, chemistry, and cybersecurity. In those cases, queries are routed through Opus 4.8 unless blocked outright for prohibited content. The problem? Some safeguards—especially in biology—have been so broad that Fable becomes nearly useless for basic questions, a point Anthropic acknowledged.

“Visible safeguards can be probed, so they have to be robust, which takes time to get right,” Anthropic wrote in a statement. “Invisible safeguards can be targeted more narrowly, allowing us to ship quickly with very few false positives. We went with invisible safeguards for this reason—and that was the wrong tradeoff. You should have visibility into the safeguards we have in place, and why. We’re sorry for not getting the balance right.”

The apology follows backlash from the AI community, which warned the secret throttling could hurt researchers trying to assess Fable’s limits. Anthropic has previously accused rivals like DeepSeek of “industrial-scale” distillation, and its terms of service already ban using Claude to build competing models.