← Back to feed

Anthropic has reversed the undisclosed policy that silently routed Claude Fable 5 requests related to frontier LLM development to Claude Opus 4.8 without informing the user. The company acknowledged the error in statements to Wired and Business Insider, with a spokesperson confirming: “We made the wrong trade-off, and we apologize for not getting the balance right.”

The restriction was documented in the Fable 5 system card, which noted that requests touching pretraining pipelines, distributed training infrastructure, and accelerator design would face “safeguards” — but the card did not disclose that those safeguards meant silent rerouting to a less capable model.

What Changed

Under the revised policy, as of this week:

• API requests that trigger the LLM development classifier now return an explicit rejection reason
• Standard user queries now visibly fall back to Claude Opus 4.8 rather than degrading silently
• Users are informed when the model switches — the same transparency that was always applied to the disclosed cyber, biology, and chemistry safeguards

The safeguards themselves remain in place. Anthropic has not removed the restrictions, only made them visible. The cyber, biology, chemistry, and model distillation routing was already disclosed in the original system card and is unchanged.

The Backlash

The stealth nature of the LLM development restriction drew immediate criticism from the ML community after researchers noticed degraded behaviour on pretraining-adjacent queries. The core complaint was not the restriction itself — competitive moat defence through safety guardrails — but the lack of disclosure. Anthropic has long positioned itself as an ethics-first, researcher-friendly alternative to OpenAI. A covert capability restriction targeting the company’s direct competitors was hard to square with that posture.

Holger Mueller of Constellation Research called the original policy “a market leader that wants to freeze the market,” a criticism Anthropic addressed implicitly by drawing a distinction between safety restrictions and competitive intent.

Scope of the Original Restriction

Anthropic said the silent fallback affected a narrow slice of requests. The LLM development classifier was tuned to catch queries specifically targeting frontier lab workflows — pretraining pipelines, training cluster infrastructure, hardware accelerator design. Standard machine learning work, coding assistance, and model fine-tuning were explicitly outside its scope.

The company said no more than 0.03% of sessions were flagged, consistent with the fraction disclosed in the original system card. What was not disclosed was that those sessions received no notice of downgrade.

The Larger Fable 5 Safeguard Architecture

The reversal applies only to the LLM development restriction. The rest of Fable 5’s tiered safety architecture is unchanged:

• Cyber, biology, and chemistry requests fall back to Claude Opus 4.8 with user notification (was disclosed at launch, remains disclosed)
• Fewer than 5% of total sessions trigger any fallback across all categories
• The full-capability Claude Mythos 5 remains restricted to vetted Glasswing partners and critical infrastructure operators
• A trusted-access programme for biology researchers is coming in the weeks ahead

The 30-day data retention requirement for Mythos-class models also stands. Anthropic confirmed that Fable 5 traffic is retained for safety monitoring and is not used for training.

What It Means for Enterprise Adoption

Enterprise buyers who paused Fable 5 evaluation over the hidden restriction have a clearer picture now. The safeguards are real and specific, but they are disclosed — which changes the risk calculus for security-sensitive deployments. Microsoft had internally restricted employee access to Fable 5 over the retention and guardrail concerns; whether the transparency change alters that posture has not been confirmed.

The episode is the first public test of whether a frontier lab can ration capability at deployment time without destroying developer trust. Anthropic’s answer — make every restriction visible — is the minimum required to stay on the right side of that line.