Detailed Analysis
Anthropic formally acknowledged in late April 2026 that a series of engineering missteps caused a measurable decline in the performance of Claude Code, its AI-powered developer tool, following weeks of mounting user complaints. The company traced the degradation to three distinct changes introduced over the preceding month, affecting Claude Code, the Claude Agent SDK, and Claude Cowork, while leaving the underlying API untouched. The most consequential of these was a March 4 change that quietly reduced the default reasoning effort level from "high" to "medium" in an attempt to address UI freezing and high latency in the Sonnet 4.6 and Opus 4.6 models — a tradeoff Anthropic later conceded was the wrong call. A second issue involved memory loss during long-running sessions, wherein Claude dropped thinking blocks, resulting in forgetfulness, repetition, cache misses, and accelerated depletion of usage limits. Anthropic confirmed fixes were fully deployed by April 20, 2026, with version 2.1.116, and added UI improvements including effort selectors and informational notices to give users more control going forward.
The handling of the episode proved as damaging to Anthropic's reputation as the technical failures themselves. User complaints about degraded performance had been circulating since at least February 2026, following the release of Opus 4.6, yet Anthropic's initial communications implied that no systemic problems existed and gestured toward user error as a likely explanation. That posture — only reversed after sustained public pressure — fueled accusations of gaslighting among a developer community that had come to rely heavily on Claude Code for production workflows. Some subscribers reported cancelling their accounts, and the belated postmortem, while technically detailed, did little to suppress speculation that Anthropic may be rationing compute resources and that the performance reductions were not entirely accidental. Anthropic explicitly denied intentionally nerfing the model and maintained that the underlying weights were never altered, but the explanations have not fully satisfied a skeptical user base.
The incident carries significant implications for Anthropic's competitive positioning in the enterprise AI tools market, where trust and reliability are paramount. The company has long sought to differentiate itself from OpenAI on the basis of transparency, safety-consciousness, and alignment with user interests — a brand identity now under strain. The sequence of events illustrates a structural tension common to AI product development: the pressure to optimize for infrastructure efficiency (reducing latency, managing server load) can create direct conflicts with the performance expectations users have come to rely on, particularly when those optimizations are implemented silently. The introduction of usage limits for subscribers alongside the fixes suggests Anthropic is actively grappling with resource constraints, lending at least circumstantial weight to the compute-rationing theories circulating in developer communities.
More broadly, the Claude Code episode reflects an emerging pattern across the AI industry in which the rapid pace of model and product iteration — often conducted without detailed changelogs or user notification — creates a trust deficit between AI companies and power users. Developer communities, which are acutely sensitive to subtle degradations in model behavior, increasingly function as informal quality-assurance networks capable of surfacing problems that internal testing misses or underweights. Anthropic's decision to publish a detailed engineering postmortem on April 23 represents a corrective step, and the company's commitment to improved code review processes and greater UI transparency signals an awareness that silent changes to reasoning defaults are no longer operationally tenable. Whether those measures will be sufficient to rebuild confidence among the developers who form Claude Code's core user base remains an open question as the competitive landscape with tools like GitHub Copilot and OpenAI's Codex-based offerings continues to intensify.
Read original article →