Some users of Anthropic's Opus 4.8 and Opus 5 models say they are spending extra time, prompts, and tokens correcting the models' confusing language, sometimes routing output through cheaper AI models to make it usable. The complaints, detailed in a GitHub issue and Reddit threads, point to a productivity and cost drag that enterprise development teams may be underestimating.
Peter Bower, founder and CEO of London-based tech startup SpaceCell, said in a GitHub issue that Opus 4.8's tendency to use confusing or invented terminology was creating extra work in software development workflows, particularly when generating code documentation. Bower wrote that he explicitly and repeatedly prompted the model to avoid certain terms and use specified alternatives, but it continued introducing unwanted terms, forcing repeated cleanup passes - including through cheaper Sonnet or Haiku models - to make the documentation "sane and presentable."
Those additional passes pushed token costs up to two times higher than they otherwise would have been, Bower said. His issue, posted last month, has received nearly 265 acknowledgements, indicating other users have faced similar problems. A ClaudeAI subreddit thread about Opus 4.8's language issues and another about Opus 5's confusing output both received significant upvotes.
Why unclear AI output slows development
For enterprise development teams, persistent language problems in the Opus models could mean real productivity losses, analysts said.
"Repeated correction cycles can erode productivity when developers spend enough time reviewing, redirecting and repairing AI output. That offsets the time saved by generating code via a coding assistant or any other tasks," said Abhishek Satapathy, principal analyst at Avasant.
That erosion also hits the operational side of the software development lifecycle. Advait Patel, senior site reliability engineer at Broadcom, said unclear AI-generated prose can affect design documentation, runbooks, architecture decision records, and incident writeups.
"A runbook written in a style that engineers find difficult or unpleasant to read, for example, could become a problem during an incident, when teams need to quickly understand and act on the information in front of them," Patel said.
Code review presents another risk. "Overly padded or confusing pull request descriptions are likely to be skimmed rather than carefully reviewed, increasing the risk of important details or potential defects being missed," Patel said.
The hidden cost of unclear output
The financial implications go beyond the monthly subscription fee. Bhupendra Chopra, chief revenue officer at IT consulting firm Kanerika, said the price for an AI coding tool does not necessarily reflect the cost of getting usable output from it. If developers make repeated passes to correct, rewrite, or review a response, or route it through another model, those steps become part of the total cost of the task, including human review time.
Most enterprises don't realize this, Patel said, because all of it "is packed into a single line item" in their coding agent bill.
That hidden cost could also affect Anthropic's ability to keep developers. Patel noted that switching coding assistants or underlying models has become relatively easy for development teams, especially as coding platforms increasingly support models from multiple providers. Enterprises may face sunk costs in configuration, hooks, and MCP setup, but the code and repos don't move, so no migration plan is needed.
"That's a genuine commercial risk for any model vendor. Low switching cost means goodwill is your only lock-in, and readability complaints erode goodwill fast because people hit them daily," Patel said.
Prompt workarounds may not be enough
Anthropic has not yet responded to Bower's GitHub issue, which also outlines changes he believes the company should make. Bower called for the model's default writing style to be closer to "a technical white paper or a good Stack Overflow answer" - "plain, declarative and direct." He also asked for the model to be less verbose and to follow instructions set in CLAUDE.md and repeated during a conversation, rather than gradually overriding them with its default communication style.
In the meantime, Patel, who said he has faced similar model drift at work with a Jenkins, Python, Terraform, GKE, and Helm stack, uses explicit rules in project configuration to ban specific phrasings when generating documentation and pull request summaries. Asking for conciseness alone can make output shorter but more cryptic, he said.
Patel cautioned that prompt-level workarounds may not be enough because model behavior can change over time. "Model behavior is a moving target. A version bump can change output register without you deploying anything, and nothing in your pipeline alerts on it."
He advised CIOs and engineering leaders to treat model behavior changes as something that needs continuous testing and monitoring. "Pin model versions for anything in a pipeline instead of tracking latest. Keep a small eval set of your own real tasks and rerun it on every model change. Track rejection and rework rate, that's your early warning. And don't let thirty teams each invent their own undocumented prompt workarounds."
Why this matters for IT and development teams
If your team uses AI coding assistants, unclear output is not a minor annoyance - it's a measurable cost in tokens, review time, and operational risk. The practical takeaway: pin model versions in your pipeline, maintain a small evaluation set of your own tasks, and track rejection and rework rates to catch language drift before it slows you down. For teams building on Claude specifically, understanding the models' quirks - and how to work around them - can make the difference between a tool that saves time and one that costs it. Training on Claude AI Courses & Certifications can help developers learn these workarounds, while broader AI for IT & Development resources address the integration challenges that come with any AI coding assistant.
Your membership also unlocks:
