Detailed Analysis
Anthropic's Claude Opus 4.7, released on April 16, 2026, is experiencing a significant service incident involving elevated error rates, specifically tied to overly aggressive Acceptable Use Policy (AUP) enforcement that is blocking legitimate user requests at scale. The incident, formally logged on Anthropic's status page on April 25, 2026, reflects a pattern of false-positive refusals that have been escalating steadily since the model's deployment. Documented cases span a wide range of benign use cases — from simple code flag modifications and computational structural biology workflows to standard technical software development conversations and prompts submitted in the Russian language, with one reported session alone generating 40 or more false positives across four interactions. The breadth of affected use cases signals a systemic miscalibration in the model's content filtering layer rather than isolated edge cases.
The incident is particularly striking given Opus 4.7's otherwise impressive capability profile. The model posted substantial benchmark gains over its predecessor, improving SWE-bench Verified performance from 80.8% to 87.6%, lifting CursorBench scores from 58% to 70%, and reducing tool errors on complex multi-step workflows by 14%. These are meaningful advances for a model positioned at the frontier of agentic and software engineering tasks. Yet the AUP refusal problem directly undermines real-world utility, creating a paradox where a technically more capable model delivers a degraded user experience. Complaints logged against AUP-related refusals have climbed from roughly two to three per month in mid-2025 to approximately eight per month by January 2026, a trend that predates but likely amplifies the current incident's visibility.
Research context suggests that Anthropic has deliberately positioned Opus 4.7 as a testbed for more stringent safety guardrails, a strategic choice that prioritizes conservative content filtering over seamless user experience. This framing contextualizes the elevated error rates not as an unintended bug but as a known tradeoff — one that has now reached a threshold where it constitutes a formal service disruption. Users have been directed to switch to Opus 4.6 or apply through Anthropic's Cyber Verification Program as interim workarounds, though neither solution is well-suited to the majority of affected users engaged in routine technical work that should not require special program enrollment.
The incident reflects a broader and unresolved tension in frontier AI deployment: how to calibrate safety filtering at a level that prevents genuine misuse without degrading legitimate utility. Anthropic's approach with Opus 4.7 appears to have erred significantly on the restrictive side, producing refusal rates that developers characterize as incompatible with normal professional workflows. This is not unique to Anthropic — OpenAI and Google have faced analogous criticism at various model release cycles — but the explicit framing of Opus 4.7 as a hypervigilant safety testbed makes the tradeoff more deliberate and thus more scrutinizable. The fact that it has now escalated to a formal status incident suggests the calibration fell outside acceptable operational bounds even by Anthropic's own standards.
The longer-term implication for Anthropic is reputational as much as technical. As Claude models are increasingly embedded in enterprise development environments, IDE integrations like Cursor, and multi-step agentic pipelines, tolerance for unpredictable refusals on benign inputs shrinks considerably. Developers building production systems require consistency and predictability, and a model that intermittently blocks routine tasks — regardless of its benchmark performance — introduces unacceptable reliability risk. Anthropic will likely need to recalibrate Opus 4.7's AUP thresholds substantially, or risk accelerating developer migration to competing models that offer more stable behavior. The incident serves as a public test of whether Anthropic can resolve the fundamental tension between its safety-first organizational identity and the practical demands of building dependable, production-grade AI infrastructure.
Read original article →