Three weeks into Opus 4.7's release, a specific group of professional users - security researchers, biology educators, virology professionals - is running into the same wall: conversations about their core subject matter are being terminated without explanation.
This is not the model refusing a sensitive request. It's classifiers - automated filters that run alongside the model and can end sessions before the model generates a response - silently resetting conversations. Topics including viral biology, cybersecurity techniques, and general biology are triggering these resets. No refusal message appears. The conversation simply stops.
Model Behavior vs. Classifier Behavior
These are different problems with different solutions, and conflating them leads to the wrong diagnosis.
Opus 4.7's trained behavior is more cautious than 4.6's. It's more likely to pause on ambiguous requests before responding, and more likely to question intent. Users report this can be managed through conversation framing - establishing professional context early in a session. That approach works on the model itself.
Classifiers are a different matter. They run before the model fully processes the conversation, which means the context you've built up in the chat does not factor into their decision. A biology researcher asking about viral replication and someone asking something genuinely dangerous reach the same filter outcome. Professional role, conversation history, and stated purpose all arrive too late.
The comparison that keeps appearing in user reports: Sonnet 4.6 handles the same biology, cybersecurity, and virology content without triggering termination. That points to a sensitivity specific to Opus 4.7, not a platform-wide policy shift.
The Practical Problem for Professional Users
Claude is used regularly by professionals who need to discuss security vulnerabilities, biological mechanisms, and dual-use topics as part of legitimate work. Security analysts need to analyze attack techniques. Biology instructors need to discuss pathogens in detail. Science writers need to ask specific questions about how things work.
For these users, the Opus 4.7 classifier behavior removes the highest-capability model from practical consideration. You cannot route around a filter that ends the session silently. The current workaround is Sonnet 4.6 - capable, but not what Anthropic positions as their most powerful model, and a real step down for tasks that benefit from Opus-level reasoning.
Anthropix has not publicly addressed the Opus 4.7 classifier sensitivity as of early May 2026. Classifier behavior is regularly adjusted after model releases, and the current thresholds may shift. Until they do, professionals working in biology or security should default to Sonnet 4.6 for daily work.