Claude Fable 5 and new AI safety fables
| Source: Interconnects (Nathan Lambert)
Tags: Claude Fable 5, Anthropic, AI safety, model restrictions, Nathan Lambert, Interconnects
Nathan Lambert's analysis: Claude Fable 5 is the strongest publicly available LLM — a remarkable benchmark leap at 2x Opus pricing — but its undisclosed safety restrictions, including silent routing of some prompts to Opus 4.8, mark the first time a frontier model has been knowingly degraded for specific users without notification.
Details
Lambert establishes the stakes first: Claude Fable 5 is definitively the strongest model available to the public, achieving a substantial jump on nearly every relevant benchmark without any single documented breakthrough technique — suggesting broad-stack advances at Anthropic. At approximately 2x Opus pricing and less than GPT-5.5 Pro, the capability-per-dollar is exceptional for non-restricted use cases.\n\nThe more consequential story is what comes bundled with the release. Anthropic implemented two tiers of safety restrictions: some disclosed to users, and some that silently modify model behavior without notification. Some prompts will be routed down to Opus 4.8 rather than Fable 5 — meaning public benchmark scores are not what all users will always receive. For frontier AI development tasks (training pipelines, ML accelerator design), effectiveness is silently degraded through prompt modification, steering vectors, or PEFT.\n\nLambert frames this as unprecedented: modern LLMs have never before been knowingly tiered for specific user categories without disclosure. The concern is not the 0.03% of developers affected today — it is the precedent and infrastructure now in place. The model was also held back 2+ months after training completion before public release, meaning a stronger successor is already in development.