Detailed Analysis
Anthropic has publicly denied that any performance degradation observed in Claude Code was the result of deliberate throttling or intentional capability reduction, instead attributing the problems to a series of discrete bugs and unintended engineering changes. User complaints, which gained significant traction on forums like Hacker News, centered on noticeably slower response times and reduced coding quality — issues serious enough that some users characterized their experiences as a form of "gaslighting" by the company. Anthropic pushed back on that framing directly, maintaining that the degradation was inadvertent and that fixes were deployed as issues were identified. The company's response represented an effort to restore trust in a product that now generates over $2.5 billion in annualized run-rate revenue, making Claude Code one of the most commercially significant AI developer tools currently on the market.
The technical root causes identified in the incident reveal the operational complexity of maintaining a large-scale AI coding assistant at production quality. One particularly damaging bug stripped "thinking tokens" from older sessions when users resumed conversations, effectively degrading Claude's reasoning capacity mid-workflow — a problem that took approximately 15 days to resolve. A separate change to prompt caching behavior in long-idle sessions silently prioritized latency over output quality with no visible indicator to users, while a system prompt modification made Claude less verbose in ways that meaningfully harmed its coding performance. That last issue was patched within four days. Taken together, these failures point to inadequate regression testing and insufficient user-facing transparency around changes that affect output behavior, rather than any coordinated effort to reduce service quality.
The Claude Code performance controversy sits alongside a broader pattern of operational stumbles at Anthropic in the same period. The company separately acknowledged accidentally leaking nearly 2,000 internal source code files due to a manual step in its release packaging process that lacked automated safeguards — an incident it described as "human error" and a "systemic process failure" rather than any individual's fault. In another unrelated incident, a company called Belo had its API access suddenly revoked due to a false positive policy violation, halting workflows for 60 employees for approximately 15 hours before access was restored. The support mechanism for that restoration was a Google Form, a detail that drew sharp criticism given the enterprise stakes involved. None of these incidents appear connected by intent, but their proximity in time underscores the scaling pressures Anthropic faces as it rapidly expands commercial deployments.
The broader context for these operational challenges is an AI industry moving faster than the institutional infrastructure supporting it. Anthropic is simultaneously managing a high-stakes commercial product, conducting frontier research, and navigating complex safety tradeoffs — as evidenced by its decision to delay public release of Claude Mythos Preview, an advanced model capable of identifying decade-old software vulnerabilities, due to concerns that cybercriminals could exploit it. That model is instead being made available only to vetted cybersecurity firms. The juxtaposition is telling: Anthropic is capable of producing genuinely powerful research capabilities while struggling with the more mundane but commercially critical task of maintaining consistent performance in production systems. For enterprise customers who have built workflows around Claude Code, the distinction between intentional degradation and accidental degradation may matter less than the fact that degradation occurred at all — and that transparency around changes remains insufficient.
The incidents collectively raise meaningful questions about the maturity of Anthropic's engineering and support operations relative to its commercial ambitions. As AI coding tools become deeply embedded in professional workflows, the tolerance for silent regressions, opaque caching changes, and limited support infrastructure will diminish. Anthropic's denial of intentionality is credible given the available evidence, and its rapid patching of identified bugs reflects a functional engineering response. However, the recurrence of distinct operational failures in a short window suggests that the company's processes for testing, communication, and customer support have not kept pace with the scale at which Claude Code is now deployed. Addressing that gap will be as important to Anthropic's commercial trajectory as any model capability improvement.
Read original article →