Claude Opus 4.8 and Opus 5 are doubling token costs for some devs
The pitch for AI coding assistants is straightforward: less work to turn intent into working software. For some users of Anthropic’s Opus 4.8 and Opus 5, the reality is running the other direction.
What Users Are Reporting
Peter Bower, founder and CEO of London-based startup SpaceCell, filed a detailed GitHub issue documenting a specific problem: Opus 4.8 keeps introducing confusing or invented terminology in code documentation, even after being explicitly and repeatedly instructed not to. The model continues using the unwanted terms despite the instructions, forcing repeated cleanup passes through cheaper Sonnet or Haiku models to make the output usable.
Those extra passes, Bower reports, are pushing total token costs up to two times higher than they would otherwise be.
The GitHub issue has collected nearly 265 acknowledgements since it was posted. A separate Reddit thread on Opus 4.8’s language coherence drew significant upvotes. A second thread focused on Opus 5 received nearly twice as many upvotes, with users describing output that is confusing and hard to parse.
The Hidden Cost Angle
Analysts and engineers quoted in coverage of the issue point to a cost accounting problem most teams miss. The price of an AI coding tool does not reflect the actual cost of getting usable output from it, according to Bhupendra Chopra, chief revenue officer at IT consulting firm Kanerika. Repeated correction passes, rerouting through secondary models, and human review time all add to the real cost of completing a task.
Advait Patel, senior SRE at Broadcom, notes that this is often invisible because it gets packed into a single line item in a team’s coding agent bill. He also flags a risk beyond documentation: unclear pull request descriptions tend to get skimmed rather than carefully reviewed, which increases the chance of defects slipping through.
Prompt Workarounds Have Limits
Bower has called on Anthropic to adjust Opus’s default writing style toward something closer to a technical white paper or a Stack Overflow answer: plain, declarative, and direct. He also wants the model to hold firmly to instructions set in CLAUDE.md rather than gradually drifting back to its default communication style. Anthropic has not yet responded to the GitHub issue.
Patel’s team uses a more targeted workaround than simply asking the model to be concise. They add explicit rules to project configuration files that ban specific phrasings, because broad conciseness instructions can produce output that is shorter but harder to understand. Even so, Patel cautions that prompt-level fixes are not a durable solution: model behavior changes with version bumps, and nothing in a standard pipeline alerts on output register shifts.
What to Do If You Are Hitting This
- Pin model versions for anything running 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 as your early warning signal.
- Ban specific unwanted phrasings in project config rather than relying on broad style instructions.
- Don’t let multiple teams build undocumented prompt workarounds independently.
The switching cost is low enough that this is a genuine retention risk for Anthropic. As Patel put it: the repos don’t move, but goodwill is the only lock-in, and readability complaints erode it fast when developers hit them every day.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
