Detailed Analysis
Claude Opus 4.7, released by Anthropic on April 16, 2026, has become the subject of an official system status incident involving elevated error rates, flagged on April 25, 2026. The incident, tracked at Anthropic's status page, reflects a confluence of two distinct but compounding technical problems: an aggressive escalation in content refusal rates driven by overly sensitive safety classifiers, and a set of breaking API changes introduced with the model that are generating 400-series errors for developers who have not yet migrated their code. The official status update, automatically surfaced to the Claude community within minutes of its posting, signals that Anthropic has acknowledged the disruption at a platform level.
The refusal rate issue appears to be the more systemic of the two problems. Anthropic used Opus 4.7 as a testbed for stricter content policy enforcement, deploying a hypervigilant Acceptable Use Policy (AUP) classifier that has generated a significant volume of false positives. Developers working across entirely unrelated domains — including computational structural biology, software development, memory authorization operations in Claude Code, and psychology applications — have reported being blocked by the classifier. One developer documented more than 40 false positives across four separate sessions. Complaint data reveals a steady escalation in refusal incidents dating back to mid-2025, accelerating through early 2026, suggesting that the guardrail tightening was a deliberate policy trajectory rather than an isolated misconfiguration. Russian-language prompts in particular appear to trigger disproportionate flagging across project types, pointing to potential bias in the classifier's training or threshold tuning.
The API breaking changes represent a separate but equally disruptive layer of the incident. Opus 4.7 introduced incompatible modifications to three core parameters — extended thinking configuration via `thinking.budget_tokens`, sampling controls including `temperature`, `top_p`, and `top_k`, and message prefilling behavior — all of which now return 400 errors when called using pre-migration code. Additionally, reasoning traces have been moved to a hidden default state, requiring developers to explicitly set visibility to "summarized" if they wish to inspect model reasoning. These changes require deliberate code migration before deployment, and teams that moved quickly to adopt Opus 4.7 without reviewing Anthropic's migration documentation would have encountered immediate breakage in production pipelines.
The incident carries broader significance in the context of Anthropic's ongoing effort to balance safety enforcement with model utility. The company has faced recurring criticism that safety-oriented guardrails, particularly in its more capable frontier models, erode practical usability for developers building legitimate applications. The Opus 4.7 situation sharpens that tension: deploying a new model as an explicit testbed for stricter content policies, while simultaneously introducing breaking API changes, created a high-friction environment for early adopters. The pattern of escalating refusal rates over the preceding nine months suggests that the current incident is not an anomaly but rather the visible peak of a longer policy enforcement curve that predates this specific model release.
More broadly, the Opus 4.7 incident illustrates a structural challenge facing frontier AI providers as their models become increasingly embedded in critical developer workflows. Breaking changes and safety classifier overfiring are not unique to Anthropic — similar tensions have surfaced across OpenAI and Google DeepMind model releases — but the scale of documented false positives in Opus 4.7's rollout, and the choice to couple aggressive safety testing with a major API revision, suggests a coordination gap between Anthropic's policy, engineering, and developer relations functions. The community's rapid mobilization around a Reddit performance megathread and the speed of the official status acknowledgment both reflect how institutionalized incident-response expectations have become among AI API consumers, and how closely the developer community now monitors the reliability of these systems as production infrastructure.
Read original article →