Recent reports of an alleged decline in the performance of Opus 4.6 have drawn criticism: the model is said to read less code, stop earlier and require more corrections on complex tasks. However, there is no direct evidence that Anthropic has deliberately weakened the model itself. It is more likely that changes in operating conditions are affecting how users perceive its quality.
Key factors include adaptive reasoning, effort settings, cache duration and usage limits. According to the documentation, Claude increasingly decides for itself when to “think” more deeply and when to handle tasks with less computational effort. This reduces costs and latency, but can sometimes lead the model to simplify complex problems too early. Shorter caching periods or changes in context compression, which may result in some loss of detail, can also affect particularly long sessions and create the impression of declining quality. For example, Anthropic has moved many Claude Code requests back to a five-minute cache, which is sufficient for shorter queries. Longer caching remains possible, but is more expensive.
The central issue therefore appears to be less about a weaker model and more about the black box of hidden policies and system optimisations. Without greater transparency around operating conditions, caching and resource allocation, users have little basis for assessing differences between sessions, and changes in those conditions can quickly look like a drop in quality.
By Ida Kandler
Source: Implicator.ai, 15 April 2026, by Marcus Schuler