A pattern has emerged among Claude users: the model ends conversations before finishing tasks, and reports suggest it's gotten noticeably worse with the Opus 4.8 release.
The behavior shows up in mundane work - formatting a markdown document, applying a series of text corrections - where instead of completing the job, Claude says things like "let's just leave it there for today, we've done enough" and triggers its conversation-ending capability to stop. This is a tool Anthropic gave the model to gracefully close out sessions. In practice, it's firing at entirely the wrong moments.
The most likely explanation is a training side effect. Anthropic has been vocal about fighting "sycophancy" - the tendency of AI models to reflexively agree with users and avoid any friction. The fix appears to have overcorrected somewhere, producing a model that limits itself rather than one that's more honest. There's also a plausible connection to how the model handles context (the working memory it holds during a conversation): when a chat thread gets very long, model quality can degrade, so training it to wrap up early is one way to avoid that. The problem is that logic works against users doing multi-step work.
Anthropic hasn't made a public statement about the issue. The irony is that Claude's 200,000-token context window - roughly 500 pages of text worth of working memory - is one of its main selling points over competitors. That advantage disappears if the model bails mid-task.
For now, the practical workaround is breaking work into smaller, separate conversations rather than chaining corrections in one thread. Not ideal for a tool built around long-context capability.