Anthropic reverses hidden frontier LLM research restrictions in Fable 5

Category: AI
Status: confirmed
Sources: Simon Willison, Anthropic Fable 5 release, HN discussion
Summary: A paragraph in the Claude Fable 5 system card disclosed that the model silently degraded responses for requests related to frontier LLM pretraining, distributed training, and ML accelerator design using steering vectors and prompt modification with no user notification. After backlash from researchers, Anthropic stated it was the wrong tradeoff and changed Fable 5 to make safeguard blocks visible. Flagged frontier-LLM requests now visibly fall back to Opus 4.8; the Messages API returns stop_reason: "refusal" with the classifier reason.
Why it matters: API integrations must handle the new explicit refusal stop reason to avoid silent result degradation on research workloads involving LLM training topics.
Follow-up: Check API behavior after Anthropic ships the visible-safeguards update; confirm no silent degradation remains.