When Users Noticed “AI Shrinkflation”
The signs weren’t imagined. Numerous posts described Claude’s decreased ability to reason through complex tasks, a tendency to repeat itself, and a sudden spike in how quickly it consumed usage credits. The problem was significant enough to unsettle much of the developer community, sparking theories and debate.
Anthropic investigated and, after an internal review, published a detailed report. Their finding: the root cause wasn’t a downgrade to the core AI model itself, but issues within the surrounding software infrastructure, or what Anthropic referred to as the “software layer.”
How Small Changes Snowballed Into Big Trouble
According to Anthropic’s report, it was a domino effect of errors that led to the headaches. The first domino fell on March 4, when Anthropic reduced Claude’s default level of reasoning effort from “high” to “medium” in order to cut response times.
The second problem struck March 26, with a major bug introduced during a cache optimization update. Instead of clearing context after an hour of inactivity, the bug wiped short-term memory with every new user interaction. This caused Claude to forget recent context, become repetitive, and forced users to restate their instructions—using up even more credits.
A third issue landed on April 16, when new system instructions meant to reduce verbosity unexpectedly cut coding task quality by about 3 percent. Together, these mishaps heightened the sense that Claude was slipping.
The reduction in reasoning made solutions feel “lazier,” and the context loss disrupted workflow. For many, the problems bred mistrust: some users speculated Anthropic was scaling back quality to manage heavy demand, a theory reportedly fueled by lack of transparency and the staggered nature of the glitches.
Outside Scrutiny and Hard Numbers
Independent analysts corroborated users’ observations. Notably, Stella Laurenzo of AMD documented a clear decline in Claude’s reasoning depth, and tests by third-party evaluators showed the model falling behind on established AI benchmarks. Because different bugs affected users at different times, the situation was confusing and hard to pin down at first.
Anthropic’s Moves to Make Amends
In response, Anthropic put a recovery plan in motion. The company immediately reset usage limits for all subscribers to compensate for the disrupted service and wasted credits. More importantly, Anthropic began overhauling its quality control. They expanded internal “dogfooding,” requiring more employees to use the exact public version of Claude, and mandated that every system prompt change pass broader, model-specific test suites before release.
Anthropic also pledged more openness. The company launched a dedicated X account, @ClaudeDevs, promising to communicate about product updates and technical issues more transparently.
Beneath the Surface: Why Even the Smartest AI Can Stumble
This episode highlights a harsh reality: top-tier AI relies not just on its underlying model but also on a fragile web of software, caches, and system configurations. When something in that web fails, even advanced systems can quickly become unpredictable or awkward to use.
The Claude incident also reflects bigger forces in the tech world. A tight “compute crunch”—the chronic shortage and high cost of GPU processing power—is driving all major AI companies, including Anthropic and OpenAI, to optimize wherever possible. In this case, efforts to speed up responses and maximize hardware efficiency backfired, undermining reliability and user trust.
The lesson for developers and users alike: pursuing efficiency and scale in AI carries real, often hidden risks. When speed is prioritized over stability, even the best-intentioned changes can spiral into public trouble that reveals just how delicate production AI systems still are.
