Anatomy of degradation: how three independent updates temporarily reduced the performance quality of Claude Code

From early March to mid-April, users of professional ecosystem tools from Anthropic, including platforms such as Claude Code, Claude Agent SDK, and the Claude Cowork collaboration system, experienced a very noticeable and systemic decline in the quality of AI performance. Engineers from the company's engineering division conducted a large-scale investigation into the incidents and reliably determined that the basic application programming interface (API) and the inference layer actually functioned without any failures. The true cause of the degradation was three independent product-level updates that accidentally overlapped in time.

The first and most significant problem arose on March 4 and was directly related to developers' attempt to balance the depth of the model's step-by-step reasoning with the responsiveness of the user interface. Immediately after the February release of the Opus 4.6 neural network in the Claude Code environment, engineers set a high default value for the reasoning effort parameter. Soon, professional users began reporting en masse that when solving complex tasks, the interface appeared completely frozen due to extremely long response times, which also led to disproportionately high token consumption. To reduce latency, the default setting was switched to a medium level.

The second technical glitch appeared on March 26 during an unsuccessful implementation of a new optimization for the prompt caching system. This algorithmic mechanism was specifically designed to radically reduce the cost of resuming work in sessions that had been inactive for more than an hour. It was initially planned that the system would remove outdated chains of intermediate reasoning once, thereby reducing the number of expensive uncached tokens sent to the API. However, due to a critical error in the code, known internally as the context clearing bug `clear_thinking_20251015`, the internal memory was erased not once after a break, but at every subsequent step of the dialogue.

The third factor that negatively affected the overall product quality was introduced on April 16, when developers added new strict instructions to the system prompt with the aim of forcibly reducing the excessive verbosity of AI responses. This particular adjustment, working in conjunction with other hidden changes to text queries, led to a completely unexpected side effect — a significant deterioration in the quality of code generation. This problem affected the widest range of current Anthropic models, including Sonnet 4.6, Opus 4.6, and Opus 4.7. Unlike the two previous glitches, the unsuccessful update of system instructions was promptly rolled back just four days later, on April 20.

The main difficulty in timely detecting this entire chain of problems was that the changes affected different, non-overlapping segments of user traffic and were implemented according to completely independent timelines. As a result, their combined effect appeared as a widespread but extremely inconsistent service degradation. The situation was seriously complicated by the fact that initially Anthropic's internal quality assessment tools (evals) and usage metrics could not reproduce the specific errors reported by clients.

This multi-layered incident clearly demonstrates the fundamental complexity of modern AI development: balancing test-time compute costs with product responsiveness. Modern neural networks are capable of solving incredibly complex tasks, but lengthy reasoning times directly conflict with users' expectations for instant GUI responsiveness and limit savings. Acknowledging that the provided experience categorically failed to meet the stated standards for professional tools, on April 23, Anthropic fully reset usage limits for all its subscribers as compensation for the technical inconveniences.