Anthropic has reversed an unannounced policy that secretly degraded the performance of its new Claude Fable 5 model, following strong backlash from developers and AI researchers. The controversy began after the release of Fable 5, a version of Anthropic’s Mythos system with national security guardrails. Anthropic had publicly said it would reroute dangerous prompts related to cybersecurity, biology, and chemistry to less advanced models, but it did not disclose that it was also silently throttling requests involving frontier LLM development.
Researchers soon discovered that when the model was used for tasks such as training competing LLMs, debugging AI code, or optimizing neural architectures, it would fail or produce weaker output without warning. Developers criticized the hidden behavior as deceptive, costly, and anti-competitive, arguing that users were paying for a deliberately crippled service. The backlash was especially sharp because Anthropic has long promoted itself as a transparent, safety-focused alternative to competitors.
In response, Anthropic admitted it had made the wrong trade-off and said it would make the safeguards visible. Under the revised policy, flagged requests will now explicitly return refusal reasons, and standard queries will visibly fall back to the older Claude Opus 4.8 model instead of being silently degraded. Anthropic defended the need for the restrictions, saying they are intended to prevent misuse of its advanced AI capabilities for dangerous research or by adversaries.
Article timeline
Build this article’s timeline
Extract dated events, key actors, and open questions using the same timeline analysis available in Signal.
Anthropic has reversed an unannounced policy that secretly degraded the performance of its new Claude Fable 5 model, following strong backlash from developers and AI researchers.
🤖 AI prefill
📅WhenWhen did this happen, or what is the timeframe?
📍Where
X
🤖 AI prefill
💡Why
Centers on Claude Fable 5, hidden AI restrictions, and frontier LLM development
🤖 AI prefill
Login to answer and earn XP for each gap you fill.
Comments
Login to view and post comments.