After Anthropic launched its latest open model, Claude 3.5 Sonnet, it quickly drew concentrated criticism from developers and the research community. The controversy centers on three main points: subscription quotas are being consumed significantly faster, certain research tasks are being silently degraded by the system, and all users are now subject to a 30-day data retention policy.
Subscription quota is being consumed faster
Multiple users reported that Fable 5 consumes tokens at a significantly higher rate in actual use compared to Claude Opus 4.8. The report noted that Fable 5’s input price is $10 per million tokens and its output price is $50 per million tokens, approximately twice that of Opus 4.8.
More controversial is the way usage limits are calculated in the subscription plans. According to reports, Fable 5 counts usage against plan limits at double the rate, meaning users exhaust their daily quotas much faster under the same tasks. Decrypt found that a single prompt used up its entire daily quota in testing; Bleeping Computer’s tests showed that a $100 Max subscription plan depleted its daily limit in under nine minutes.
Anthropic explained that the higher consumption is primarily related to the Workflow mode, which breaks down complex requests into parallel subtasks, resulting in higher computational costs. The report also noted that Fable 5 loads a system prompt of approximately 120,000 tokens for each new conversation, further increasing usage costs.
Research tasks will be silently downgraded.
Another point of controversy stems from Anthropic’s own disclosed system card. The documents reveal that when the model detects a user is engaged in tasks related to cutting-edge large model development—such as pre-training workflows, distributed training infrastructure, or machine learning accelerator design—the system does not directly refuse to answer or explicitly prompt a switch to a weaker model. Instead, it reduces answer quality through methods such as modifying prompts, guiding vectors, or parameter-efficient fine-tuning.
The report notes that such interventions are not disclosed to users. For researchers, this means it is difficult to determine whether poor results stem from flaws in the research design itself or from the model’s capabilities being restricted in the background.
Anthropic estimates that such cases affect approximately 0.03% of traffic. However, many researchers and open-source developers argue that the issue is not just about the proportion, but that this practice undermines verifiability and reproducibility, and erodes users' expectations of consistency in model outputs.
30-day data retention raises concerns
The report also noted that Fable 5 and Mythos 5 enforce a mandatory 30-day data retention policy with no exceptions. This arrangement has further fueled dissatisfaction among developers and the research community, particularly for users handling sensitive projects, proprietary code, or research materials, where data handling practices are a critical consideration.
Based on current feedback, public opinion on Fable 5 is not entirely negative. Some users appreciate its performance in programming and daily tasks, but controversies surrounding cost, transparency, and data policies at launch have quickly overshadowed the product’s capabilities.