Forcing Claude into JSON mode often doubles your bill without improving results.
A small SaaS team in the EU (case-002) saw monthly costs drop from $1,840 to $287 when they stopped paying for retry loops on malformed JSON. The outputs were identical—confirmed by a blind A/B test on 50 samples—but the forced structure triggered repeated attempts that billed as separate calls. Their support auto-reply prompts were high-volume and repetitive, making the waste impossible to ignore.
An agency in the UK (case-004) running agentic workflows faced the same issue. Their research-draft-critique loops cost $2,490 before switching, then $498 after. The problem wasn’t the JSON itself, but the retries when the model failed to conform. The savings compounded because each loop iteration could trigger multiple attempts.
Even everyday coding tasks aren’t immune. A developer (case-006) using Claude as a daily assistant saw bills fall from $145 to $29 by avoiding forced-JSON retries. The fix didn’t change the quality—just the cost.
Audit your own Claude usage. Paste your last 30 days at aiusage.ai — no signup for the number.