2 days ago
25
Anthropic’s newest AI model was live for roughly 24 hours before users figured out something was off. Claude Fable 5, the company’s first Mythos-class model available to the general public, had been quietly rerouting certain queries to a less capable model without telling anyone. When the AI community noticed the performance drops and called it out, Anthropic did something rare in tech: it admitted the mistake.
The company has now promised visible safeguards going forward, meaning users will actually know when their queries are being flagged or redirected. The catch? Anthropic warned that fixing the transparency problem will come with a side effect: more false positives as the company recalibrates its classifiers. In English: expect the system to occasionally flag perfectly innocent questions while it learns to better distinguish between genuine risk and a biology student doing homework.
What actually happened
Claude Fable 5 launched on June 9, 2026, marketed as a major leap in Anthropic’s model lineup. Under the hood, the Mythos-class architecture included a new safety layer designed to handle high-risk queries, particularly in sensitive domains like cybersecurity and biology.
When the system detected a potentially dangerous prompt, it would silently redirect the conversation to Claude Opus 4.8, an older, less capable model. No notification, no explanation, no opt-out. The user just got worse answers and had no idea why.
Anthropic says this fallback mechanism triggered in fewer than 5% of user sessions. That sounds small until you consider the scale of Claude’s user base and the fact that many of the affected users were likely power users probing the model’s capabilities in technical domains, exactly the audience most likely to notice a sudden drop in quality.
The backlash was swift and pointed. Users accused Anthropic of what amounted to invisible performance sabotage, a term that spread quickly through developer forums and social media. The core complaint wasn’t really about safety. It was about the secrecy. Developers and researchers tend to accept reasonable guardrails when they’re clearly communicated. Hiding them is a different story entirely.
The transparency pivot and its costs
Anthropic’s response came within a day of the controversy gaining traction. The company acknowledged the approach was a miscalculation and committed to making all safety-related redirections visible to users going forward.
But transparency is not a free upgrade. Anthropic explicitly cautioned that the shift would lead to more false positives as it refines its classifiers.
Beyond the classifier changes, Anthropic introduced a mandatory 30-day data retention policy for all Mythos-class models. This applies universally, with no opt-out even for enterprise partners. Enterprise customers who handle sensitive data now have to factor in a minimum retention window they cannot negotiate around.
There is one carve-out, though. Anthropic plans to offer non-restricted versions of its Mythos-class models to vetted partners in the life sciences sector. The logic is straightforward: biomedical researchers need access to the full model capabilities without safety redirections interfering with legitimate work on topics that happen to overlap with high-risk domains.
Why this matters beyond AI drama
Anthropic’s competitors are watching closely. OpenAI, Google DeepMind, and others building comparable frontier systems will inevitably face similar decisions about whether to silently degrade outputs or openly flag content restrictions. Anthropic just demonstrated that the silent approach generates backlash fast enough to force a policy reversal within 24 hours.
The 30-day retention policy is perhaps the sleeper issue here. Enterprise customers evaluating Anthropic’s Mythos-class models now have to weigh a non-negotiable data handling requirement against whatever performance advantages the new architecture offers. For companies in regulated industries like healthcare or finance, that mandatory retention window could create compliance conflicts with their own data governance frameworks, potentially pushing some enterprise buyers toward competitors willing to offer more flexible terms.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
English (US) ·