Claude Opus 4.7 refusal rate surges to 30+ monthly complaints as Anthropic tests aggressive guardrails

Anthropic's Claude Opus 4.7 is blocking legitimate developer requests at an unprecedented rate following the deployment of more aggressive safety filters in April 2026. Developers filed more than 30 complaints about false positive refusals in April, compared to 2-3 monthly reports from July through September 2025.

The numbers

According to a graph compiled from Claude Code GitHub issues:

• July-September 2025: 2-3 AUP-related complaints per month
• October-November 2025: 5-7 complaints per month
• January-March 2026: ~8 complaints per month
• April 2026: 30+ complaints

Anthropic stated it is "releasing Opus 4.7 with safeguards that automatically detect and block requests that indicate prohibited or high-risk cybersecurity uses" to prepare for the eventual release of Mythos, a model the company claims is too capable at vulnerability discovery to release publicly.

What's being blocked

Developers report Claude Opus 4.7 refusing:

• Computational structural biology tasks (Issue #49751) that worked in version 4.6
• Reading a PDF of a Hasbro Shrek toy advertisement (Issue #48723), with the specific trigger identified as the PDF text "CHARACTER OR FOR DONKEY UNDERNEATH"
• Proofreading cybersecurity lab exercises for a textbook by LSU Cyber Center director Golden G Richard III (Issue #50916)
• Processing Russian-language prompts across unrelated projects including psychology books and web apps (Issue #48442)

Security researchers with approved Cyber Use Case Exemptions report the bypass mechanism doesn't work via API (Issue #49679), despite functioning in the web interface.

Anthropic's position

Anthropic has not publicly disclosed pricing changes or technical specifications for the Opus 4.7 guardrails. The company stated: "What we learn from the real-world deployment of these safeguards will help us work towards our eventual goal of a broad release of Mythos-class models."

What this means

Anthropic appears to be using paying customers as test subjects for guardrails designed for a future model, with refusal rates increasing 10-15x in six months. The company's own exemption system for legitimate security research doesn't function properly via API. Developers paying $200+ monthly for Pro access are effectively beta testing safety filters at the expense of productivity, with no indication when or if the false positive rate will normalize.