Users Say Opus 5 is Too Verbose (And Getting Worse)

A wave of complaints has followed Claude Opus 5 since launch, and unusually, Anthropic's own team has been publicly agreeing with most of them. On X, developer Anton Ivanov summed up the split many users describe: strong on tough reasoning and science, but occasionally "lazy or stop." Another user, Eve, said the writing quality dropped enough that she switches back to the prior version, Claude 4.6, "all the time." When pressed on the pattern, Anthropic's Thariq Shaukat replied directly: "Yeah I agree that Opus 5 is a really spiky model and we want our models to be consistent and warm and feel like Claude. We're working hard on this and it is a huge priority for us." OrcaRouter

The gap between the scoreboard and the chat window

The strange part is that Opus 5 tests well. It scored competitively against GPT-5.6 on coding, agentic search, and computer-use benchmarks at launch. The complaints are almost entirely about the daily-use experience: excessive explanation, over-engineering small requests, and a conversational tone users describe as denser and less warm than earlier Claude models. One coding-tool figure called it the most annoying model he'd used, saying it treats minor issues like major ones requiring large rewrites. Anthropic itself supplied the likely explanation: Thariq separately disclosed that the team stripped more than 80% of Claude Code's system prompt for Opus-class models after finding the old rules contradicted each other, with no measurable drop on coding evals. Fewer guardrails plausibly means more of the model's raw instincts showing through unfiltered, in both directions.

Not every complaint is really about the model. Some users report their weekly usage draining at wildly different rates for what feel like identical tasks, and others have hit account or billing snags unrelated to model quality. Those are real friction points, but worth separating from the verbosity and tone complaints, since lumping every gripe into "Opus 5 is bad" makes it harder for anyone, including Anthropic, to fix the actual problem.

What this means if your team runs on Claude

Anthropic's response here is the part worth noting for anyone building a vendor relationship on AI tools: public acknowledgment, a named cause, and a stated priority, rather than silence or deflection. That's a healthier posture than most software vendors manage. It doesn't mean the current experience is acceptable for your workflow today. If verbosity or inconsistency is costing your team real time, the fix right now is prompting discipline and model selection, not waiting on a patch.

If you're trying to figure out which Claude model actually fits your team's workflow this week, not which one has the best benchmark, that's exactly the kind of practical call our growth strategy work is built around, and our AI marketing services team can help you build prompts that route around the rough edges.