The classifiers Anthropic puts in front of Fable are too zealous (combine-lab.github.io)

Anthropic's recent release of the Fable model—a "safety-conscious" version of its predecessor Mythos—has stirred significant controversy in the AI/ML community. Initially launched on June 9, 2026, it faced immediate export controls just days later, preventing access to anyone outside the U.S., including Anthropic employees. Following a brief negotiation period, access was restored, but not without notable drama online concerning the model’s utility, particularly for research-level tasks. Many users, including researchers in the field of biology, reported their prompts being rejected due to overly conservative safety measures that seem to misinterpret benign requests as potentially risky. A critical examination of Fable was provided by Rob Patro, who attempted to leverage the model for porting his widely-used RNA-seq tool, Salmon, to Rust. However, Fable flagged and rejected his inquiries based on misunderstood biological terminology, highlighting a pervasive issue with its classifier system. Patro also explored the model's ability to assist in abstract computational problems related to network evolution—another avenue where Fable failed to engage, clinging to its overly cautious safety protocols. This experience has raised questions about the model's appropriateness for tasks that necessitate nuanced understanding, indicating that despite Fable's design for safety, its effectiveness is severely hampered by misaligned filtering mechanisms, ultimately limiting its use in scientific exploration.