Claude Code Opus 4.7: 16B cache reads across 8 sessions, forensic JSONL data (github.com)

Claude Code has recently faced significant issues with inflated session usage and abnormal rate limits, primarily impacting users since March 22, 2026. Users have reported that their experiences with the Claude Model Opus have drastically changed, leading to faster exhaustion of session limits under previously moderate usage patterns. One user detailed a 334,603 quota-pressure estimate from standard usage, which is problematic as it suggests excessive load for minimal input and output. This has triggered concerns that the Opus model may be encountering performance regressions since earlier versions, ultimately degrading user experience and quality of output. For the AI/ML community, this situation highlights the challenges in maintaining consistency and reliability when rolling out new features or versions. The reported issues suggest that the model's architecture or recent updates, including cache read efficiency and session management, may not handle user demands effectively. As developers and researchers strive to optimize AI systems, these discussions serve as crucial reminders of the importance of performance stability, user feedback, and continuous monitoring in AI deployments.