By early April, developers relying on Claude Code for daily work noticed something off. Sessions that once held deep context started resetting mid-task. The model would claim refactors were complete when files remained unchanged. Hallucinations crept back. Complaints on forums escalated from scattered gripes to widespread frustration.

Anthropic confirmed the degradation on April 23. Three separate engineering changes — none involving model weight updates — combined to erode performance. The API stayed stable, but the agentic coding tool suffered.

First, on March 4, engineers lowered default reasoning effort to fix interface lag. The trade-off cut thinking depth significantly. It was reversed on April 7. Second, a caching bug introduced around March 26 wiped short-term memory in longer sessions, making outputs inconsistent. Third, system prompt adjustments alongside a new Opus release caused a measurable 3% performance drop on coding benchmarks. That change was pulled on April 20.

All three issues were fixed in version 2.1.116 by April 20. Anthropic reset usage limits for affected subscribers. But the episode reveals a deeper engineering challenge: small backend tweaks — made for speed or efficiency — can cascade across user experience when not tested against real workloads.

Users detected the shift before the company did. One analysis on GitHub logged thinking depth dropping roughly 67% across thousands of sessions. Stella Laurenzo at AMD examined session logs and found faster, shallower reasoning favoring superficial fixes over sustained analysis. The community built an evidence chain through ablation studies and before-and-after comparisons.

Anthropic initially dismissed reports, suggesting prompt refinement or rising expectations. Only after external pressure mounted did the company produce a detailed postmortem. The delay damaged credibility. Developers don't treat coding agents like chatbots. They integrate them into pipelines. A tool that lies about completing work or drops context after 10 turns breaks trust fast.

The company pledged better monitoring, longer ablation tests, and clearer communication. For now, many engineers verify outputs against older baselines and approach each deployment with extra scrutiny. The lesson extends beyond one lab: frontier models operate in a narrow band where small changes in effort, context handling, or prompting produce outsized effects. Users now detect those shifts faster than builders sometimes expect.

Source: Webpronews