Anthropic Cut One Customer's Claude Costs 80% Without Touching Accuracy

One Anthropic customer reached 90.5% accuracy on its task at one-fifth of its original Claude spend — the standout number in cost-optimization guidance and tooling Anthropic published September 8 for the Claude Platform. The post organizes savings around three levers: raising prompt-cache hit rates, eliminating outdated prompting habits carried over from older, less-capable models, and calibrating “effort” level to task complexity rather than defaulting to maximum. Anthropic backs each lever with benchmark numbers: a prompt-pattern audit on a customer-support benchmark cut cost 14.6% while accuracy actually improved 5.3%; caching plus batch processing cut LegalBench costs roughly 58%; prompt caching alone dropped tau2-bench retail costs about 73%; and OfficeQA Pro fell from $136.20 to $64.87 per run. Anthropic named the anti-patterns responsible for the waste explicitly — “verification rituals,” “thoroughness boosters,” mandatory procedures, and stale few-shot examples inherited from earlier prompting eras — habits that cost tokens on models now tuned to work well from direct instruction.

The tooling matters as much as the numbers: Anthropic shipped three CLI-style commands — /claude-api prompt-audit to detect the anti-patterns above, /claude-api cost-optimize to profile token spend and propose cuts, and /claude-api hillclimb for iterative tuning — treating cost governance as a platform feature rather than something customers have to reverse-engineer themselves.

That instinct to re-benchmark rather than assume matches what Rippling found when it tested 15 models against 2,100 scored runs on real production payroll data: accuracy flattened to within a single point across the top seven models while price varied three to seven times, and GLM 5.2 matched Opus 4.6’s accuracy tier at roughly 43% of the cost. Anthropic optimizing its own platform’s cost curve and Rippling finding wide price variance for comparable accuracy point at the same operating discipline for any firm running production AI workloads: cost is now a variable to actively manage and re-test against each model release, not a fixed line item to accept.