Claude Opus 5 and GPT-5.6 Sol, the latest models from Claude and ChatGPT, are finding excessive bugs and proposing elaborate fixes, burning through users' token allowances. A PCWorld senior writer, using a Claude Pro and ChatGPT Plus subscription ($20/month each), reported blowing half their weekly allowance chasing issues found by the models. The nitpicky behavior affects not just coders but anyone editing proposals or reviewing spreadsheets on a token budget. The writer asked Claude Opus 5 for an "on a budget" prompt, which produced a six-point plan; GPT-5.6 Sol reviewed it with a dozen criticisms, then backtracked and narrowed the list to three. The final agreed prompt, pared down to four bullet points, instructs the AI to treat time and usage budget as a real constraint, match effort to task worth, apply the smallest fix that fully solves the problem without scope creep, separate what matters from what doesn't, and rank by value-for-cost with rough effort (S/M/L) per item.