Detailed Analysis
A Claude Code user on the $100/month Claude Max subscription plan has reported a dramatic and unexplained performance degradation over the past two weeks, with routine bug-fixing tasks that previously resolved in two to three minutes now consuming twenty to thirty minutes and burning significantly more tokens. The user, working with a Node.js backend stack, emphasizes that no changes were made to their context window configuration, system prompts, or usage patterns, suggesting the slowdown originates on Anthropic's infrastructure side rather than from any user-introduced variable. The degradation is described as consistent and reproducible across similar small, localized bug fixes that had previously been handled efficiently.
The issue reflects a pattern that has emerged across AI coding assistant platforms as user bases scale rapidly. When a model or its supporting infrastructure experiences increased demand — whether from a surge in subscribers, backend model updates, or changes to rate-limiting and resource allocation policies — latency and token consumption can increase even when the user-facing plan tier and pricing appear unchanged. The Claude Max plan is Anthropic's premium consumer subscription tier, which promises higher usage limits than the standard Pro plan, but does not necessarily guarantee dedicated compute or consistent response latency at the infrastructure level. Changes to how Claude internally plans and executes multi-step agentic tasks, such as those performed during code editing, could also contribute to expanded token usage without visible prompt changes by the user.
This situation connects to a broader challenge in productizing large language model agents for developer workflows. Tools like Claude Code, GitHub Copilot, and Cursor operate in an environment where the underlying model is continuously updated, sometimes silently, and where infrastructure load is highly variable. Users building workflows around consistent performance benchmarks — such as "this task takes three minutes" — are inherently vulnerable to regressions introduced through model updates, system prompt changes baked into the product layer, or changes to agentic scaffolding like planning loops and tool-call chains. Anthropic has been expanding Claude Code's agentic capabilities significantly, and more sophisticated internal reasoning or extended thinking behavior, if toggled on or adjusted, could explain both the time increase and the token inflation observed.
From a product trust perspective, the opacity of these changes is a recurring friction point for power users. The poster's frustration stems not just from the slowdown itself but from the absence of any changelog or communication explaining the shift in behavior. Anthropic, like most AI providers, does not routinely publish granular release notes for infrastructure or model behavior changes affecting tools like Claude Code. This leaves developers in a diagnostic vacuum, unable to distinguish between a billing-tier throttle, a model update, a regional infrastructure issue, or a bug in the agentic scaffolding. Community forums and platforms like Reddit have consequently become de facto debugging spaces for these kinds of regressions, with users pooling observations to identify patterns that vendors have not formally acknowledged.
The practical remediation options available to affected users are limited but worth examining. Reducing context window size, switching to non-agentic single-turn prompting for simpler fixes, or explicitly instructing Claude to minimize planning steps before acting are common community-suggested workarounds that can reduce token overhead. Monitoring whether the slowdown is model-specific — for example, whether it occurs identically across Claude 3.5 Sonnet and Claude 3.7 Sonnet within the Code interface — can help isolate whether the issue is model-layer or orchestration-layer. However, absent transparency from Anthropic about what changed in the two-week window the user identified, definitive diagnosis remains elusive, underscoring the broader need for AI product vendors to provide better observability tooling and behavioral change notifications to their developer user base.
Read original article →