Anthropic admitted this week that invisible safeguards in its Claude Fable 5 model were 'the wrong tradeoff' and announced it will replace them with visible fallbacks to Claude Opus 4.8, starting this week. The company faced backlash after launching Claude Fable 5, the first of its new Mythos class, with a safeguard buried in its 319-page system card that secretly degraded responses for users suspected of building competing AI models. The controversy erupted after AI research firm SemiAnalysis publicly reported on June 9, 2026 that their GPU inference research had been flagged, and Anthropic posted an apology on X on June 11, 2026. The invisible safeguard worked differently from the model's existing visible protections for cybersecurity and biology research, which notified users when requests were rerouted to the older Opus 4.8 model.
Anthropic Announces Visible Fallback System for Flagged Requests
Starting this week, flagged requests will visibly route to Claude Opus 4.8 instead of silently delivering degraded Fable output. API users will receive a stated reason when a request gets refused. Anthropic said server-side fallback notifications will roll out in the next few days. The company posted on X: '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.'
Claude Fable 5 Originally Used Silent Response Degradation
The LLM-development safeguard detected when users were working on pretraining AI systems, building distributed training infrastructure, or designing machine learning chips. The model would silently alter its own behavior through prompt modification, steering vectors, or parameter tweaks to give a worse answer without notification. Users received a response but not from the Fable 5 they paid for. Claude Fable 5 already had visible safeguards for cybersecurity and biology research that notified users when requests were rerouted to the older Opus 4.8 model. The classifier's precision issues led to legitimate machine learning work getting flagged, creating reproducibility problems for AI researchers who had no way to know their results were contaminated.
New System Routes Flagged Requests to Claude Opus 4.8
Flagged requests will now visibly fall back to Opus 4.8, the same as the company's safeguards for cyber and bio research. Users will see this notification every time it happens. On the API, any flagged request will return a reason for refusal rather than silently delivering a degraded answer. Anthropic is applying the same changes to its biology and cybersecurity classifiers, which had drawn complaints about flagging harmless research prompts.
Anthropic Acknowledges Increased False Positives from Visible Safeguards
Anthropic directly admitted the tradeoff it is accepting: making safeguards visible makes them easier to bypass, which means the classifier has to cast a wider net to remain effective. More false positives—legitimate machine-learning work that gets caught and rerouted—are coming while the company tunes its systems. Anthropic said it is working to reduce false positives 'as fast as possible' but offered no timeline. Fable 5 remains free on Pro, Max, Team, and Enterprise plans until June 22, after which it shifts to API usage credits only.
FAQ
What did Anthropic change about Claude Fable 5's safeguards this week?
Anthropic announced starting this week that flagged requests will visibly route to Claude Opus 4.8 instead of silently delivering degraded output. API users will receive a stated reason when requests get refused, and server-side fallback notifications will roll out in the next few days.
Why did Anthropic apologize for Claude Fable 5's original safeguards?
Anthropic apologized because the model's invisible safeguards for LLM development secretly degraded responses without user notification, which the company admitted was 'the wrong tradeoff.' The safeguard was buried in a 319-page system card and caused reproducibility problems for legitimate AI researchers who had no way to know their results were contaminated.
When does free access to Claude Fable 5 end?
Fable 5 remains free on Pro, Max, Team, and Enterprise plans until June 22, after which it shifts to API usage credits only.
