Anthropic Opus 4.8, 5 Drive AI Coding Costs Up to 2x as Users Rework Output

3 articles · Updated · InfoWorld · Aug 20

Up to 2x higher token costs are hitting some Anthropic Opus 4.8 and 5 users, who say confusing, invented or overly verbose language forces repeated cleanup and extra prompts.
Peter Bower, CEO of SpaceCell, said Opus 4.8 kept ignoring explicit wording instructions in documentation tasks, pushing his team to run outputs through cheaper Sonnet or Haiku models to make them usable.
Nearly 265 GitHub acknowledgements and active Reddit threads suggest the issue is broader, with users also reporting similar hard-to-parse output from Opus 5.
Analysts and engineers said the problem can erode productivity across documentation, runbooks, ADRs and code reviews, while hidden rework time inflates the real cost of AI coding tools.
Anthropic has not responded to Bower’s issue; engineers say prompt bans can help, but enterprises should pin model versions and continuously test behavior because output drift can change without warning.