Anthropic’s unveiling of its most powerful public model on Tuesday was met with a swift and significant backlash from the artificial intelligence community by Wednesday.
Claude Fable 5, the first publicly available iteration of the company’s restricted Mythos-class technology, has garnered praise for its coding prowess and impressive performance in general use. However, its launch has been marred by widespread criticism regarding its exorbitant token consumption, covert self-sabotage for specific research tasks, and a mandatory 30-day data retention policy for all users.
The negative reaction was immediate and vocal, resonating across researchers, developers, founders, and open-source advocates, signaling more than just typical launch-day complaints.
Costly Token Consumption
Users’ initial concerns centered on Fable 5’s pricing structure. The model charges $10 per million input tokens and $50 per million output tokens, a doubling of the rates for Claude Opus 4.8.
This aggressive pricing is compounded by its impact on subscription plans. Fable 5 counts double against usage limits compared to Opus, meaning users deplete their plan allowances twice as quickly. Early tests revealed extreme token consumption, with one user reporting their daily quota was exhausted by a single prompt. Bleeping Computer’s testing indicated that a $100 Max subscription’s daily allowance was consumed in under nine minutes.
Scrimba CEO Per Borgen calculated that Fable 5 burned 1.3 million tokens in seven minutes, equating to an hourly cost of $160, or an annual salary equivalent of $333,000. Other users reported spending over $1,000 in tokens in a single day on a $200 subscription plan, with one CTO lamenting the inability to conduct meaningful testing due to rapid quota depletion.
Anthropic attributes the high token usage in “Workflow mode” to its design of breaking down complex prompts into parallel subagent tasks, which inherently requires more compute. Additionally, a new system prompt, approximately 120,000 tokens long, is loaded into every new conversation. The company maintains that Fable 5’s per-task efficiency is superior, producing more comprehensive results with fewer iterations, a claim that has been met with skepticism by users experiencing rapid budget depletion on live subscriptions.
Covert Self-Sabotage for Research
A more damaging criticism emerged from Anthropic’s own documentation, which revealed that Fable 5 silently nerfs itself when it detects users are engaged in frontier large-language-model development. Instead of refusing to respond or falling back to a less capable model, it modifies prompts, steering vectors, or employs parameter-efficient fine-tuning without user notification.
This covert degradation means researchers cannot be certain if they are receiving full Fable 5 responses or diminished Opus-level output. Anthropic stated in Fable’s System Card that these safeguards would not be visible to the user, unlike interventions for cybersecurity or distillation attempts.
This distinction is critical for researchers. While an open refusal or a detectable fallback to a weaker model allows for understanding boundaries, a model that appears to assist while secretly delivering inferior results undermines scientific reproducibility. A failed outcome could be misattributed to the researcher’s methodology rather than an undisclosed model intervention.
Anthropic estimated this would affect a mere 0.03% of traffic, a figure dismissed by the open-source and research communities as irrelevant to the principle at stake. Arthur Zucker, a core contributor at Hugging Face, expressed a loss of trust, stating, “My tokens will no longer fly your way.”
Mikel Artetxe, cofounder of Reka AI, drew parallels to hypothetical scenarios where companies would sabotage competing technologies. He argued that the researchers most affected were academics, startups, and independent developers—the very individuals Anthropic’s safety rhetoric claims to protect. AlphaXiv, an open research platform, condemned the practice as a dangerous precedent, emphasizing that safety policies must be transparent and auditable.
Nathan Lambert, formerly of the Allen Institute, characterized Anthropic’s actions as “anti science, and therefore anti progress and anti safety.” Pseudonymous user “CalleBTC,” a developer awaiting Fable to train a world model, accused Anthropic of stifling competition and acting unethically.
Concerns were also raised that Fable’s restrictions might extend beyond specific topics, potentially influenced by how the model classifies users. One user reported being unable to interact with Fable 5 without disabling memory features, as the model identified them as a biomedical researcher.
Mandatory Data Retention Policy
The third major grievance directly impacts enterprise users but carries broader implications. Anthropic’s announcement mandates 30-day data retention for all traffic on Mythos-class models, including Fable 5, across all platforms, including third-party services like AWS Bedrock and Google Vertex AI. While Anthropic assures data deletion after 30 days in “almost all cases,” the policy poses significant challenges for companies handling sensitive information such as legal communications, healthcare records, or confidential source code. Users argue that existing privacy agreements need to be updated to guarantee data protection.
The data retention policy also creates compliance issues for European companies operating under GDPR’s data minimization rules. Lisan al Gaib, a prominent figure in the AI community, highlighted that companies requiring zero data retention are effectively excluded from using Fable 5.
Hugging Face CEO Clement Delangue framed these events within a larger concern about the concentration of power and wealth in AI, advocating for open science and open-source development. Another user criticized Anthropic’s CEO for allegedly being against the technology he creates, restricting knowledge and education in machine learning.
Fable 5 is currently available for free on Pro, Max, Team, and Enterprise plans until June 22. Following this period, it will transition to usage credits only, with Anthropic stating broader access will be restored as capacity expands.
