A growing number of Claude users are reporting that Opus 4.6, Anthropic's flagship model, has started giving instructions instead of doing the work. The complaints follow a familiar pattern: the model suggests steps, outlines approaches, or tells users what to do rather than writing the code, drafting the text, or completing the task directly.
This is a recurring phenomenon across large language models. OpenAI users reported similar behavior with GPT-4 during late 2023, and the pattern tends to surface every few months with various models. The cause is never fully clear. It could be subtle changes in system prompts, shifts in how the model is served, or even confirmation bias where users notice unhelpful responses more after reading about them.
For practical purposes, there are a few things worth trying. Being explicit in your prompts helps: "Write the complete code" works better than "Help me with this code." Adding instructions like "Do not explain what to do, just do it" can reduce the delegation behavior. Some users find that starting a fresh conversation produces better results than continuing long threads where the model may have settled into an advisory tone.
Anthropic has not publicly acknowledged any recent changes to Opus 4.6's behavior. The company does adjust model outputs through system prompt updates and RLHF tuning (reinforcement learning from human feedback, where the model is trained on human preferences), which can shift response patterns without a formal model version change. Without version-level transparency, users have no way to confirm whether the model they are using today behaves identically to the one they used last week.
The frustration is real, but so is the ambiguity. Until AI companies provide changelogs for behavioral updates the way software companies do for code, these reports will remain difficult to verify or resolve.