Detailed Analysis
Users on Anthropic's Max 20X subscription tier are reporting a notable degradation in the usability of Claude Cowork following the release of Opus 5, specifically around aggressive and frequent "compacting" of conversations. According to the original poster, this compaction—Anthropic's mechanism for summarizing or trimming conversation history to manage context window usage—is now triggering every few prompts, even in relatively fresh sessions where token usage remains well under 200K. This is notable because Max 20X subscribers are ostensibly entitled to a 1 million token context window, making the frequency of these interruptions disproportionate to the actual context load being generated. The core question raised—whether Anthropic quietly adjusted compaction thresholds coinciding with the Opus 5 rollout—points to a broader pattern of opaque backend changes that accompany major model releases.
This kind of complaint matters because context window management sits at the heart of what makes agentic coding and collaboration tools like Cowork useful in practice. Cowork is designed for extended, multi-turn technical work where maintaining continuity of context—file states, prior decisions, accumulated reasoning—is essential to productivity. When compaction fires too often, users lose fidelity in the conversation, forcing them to re-establish context, repeat instructions, or watch the model "forget" earlier decisions. For paying customers on a premium tier specifically marketed around expanded context capacity, this undermines the core value proposition of the subscription. It also erodes trust: users on forums like Reddit frequently interpret undisclosed changes to context handling as a form of quiet cost-cutting, since maintaining full context windows is computationally expensive, and providers have incentive to trigger summarization earlier than users expect, especially at scale.
This incident fits into a recurring tension in the AI industry between advertised capabilities and real-world enforcement, particularly around context windows, rate limits, and usage caps. Similar controversies have played out with other providers—OpenAI, Google, and others have all faced user backlash after seemingly reducing effective limits without clear communication, often attributed to infrastructure load balancing following high-profile model launches. Opus 5's release likely brought a surge in demand for Cowork specifically, and it's plausible Anthropic adjusted compaction heuristics to manage server-side compute costs or latency, even if the advertised context window figures remained officially unchanged. Whether this was a deliberate policy shift, an unannounced backend optimization, or a bug introduced during the Opus 5 rollout is unclear from the available information, and Anthropic has not issued public comment on the specific behavior described.
More broadly, this reflects the growing pains of shipping frontier models into production tools used by paying professionals who rely on consistent, predictable behavior for real work. As agentic and "cowork"-style products become more central to Anthropic's business strategy—competing directly with tools like GitHub Copilot Workspace, Cursor, and OpenAI's Codex-style offerings—maintaining transparency around resource allocation, context handling, and any silent throttling becomes increasingly important to retaining developer trust. Community-sourced bug reports like this one often serve as an early warning system, surfacing discrepancies between marketed capabilities and shipped behavior faster than formal support channels, and they tend to pressure companies toward public acknowledgment or fixes, particularly when tied to premium-tier promises like the 1M token context window.
Read original article →