One Command Cut Claude's Cost by Up to 73%!

Hey everyone, it's Shii! Today I want to share something I found on the Claude Blog. It's about how to cut the cost of apps built on Claude Platform without giving up performance, and I thought it was such a practical, interesting read that I had to share it right away!

Claude Blog claude.com

What was announced?

What Claude published on its blog is an article about how to reduce the cost of Claude Platform apps without sacrificing performance. Cost and performance are often assumed to be a trade-off, but this article shows that three fixes can often cut cost while keeping performance intact:

• Maximize your prompt cache hit rate
• Remove prompt "anti-patterns" when upgrading to a frontier model
• Calibrate effort (how deeply the model deliberates) to the task

All three pieces of guidance are packaged into a skill called claude-api, which you can use directly from inside Claude Code.

Why it matters

Autonomous agents like Claude Code use more tokens the longer a conversation runs. And every time a new model comes out, old prompts tend to keep instructions that were really just patches for an earlier model's weaknesses, and those can quietly hurt both cost and accuracy. According to the article, common "anti-patterns" look like this:

• Verification rituals ("double-check your work," "verify twice before responding")
• Thoroughness and emphasis boosters ("be maximally thorough," "CRITICAL: YOU MUST ALWAYS...")
• Mandatory procedures and scratchpad scaffolds (templates like "think step by step in a scratchpad")
• Stale few-shot examples tuned to an older model's failure modes
• Contradictory rules ("always refund within policy" vs. "never issue refunds without escalation")

Frontier models tend to follow instructions more literally, so these contradictions and excess instructions can trigger extra tool calls and wasted deliberation, hurting cost and accuracy at the same time. That's exactly why the article's core message is: revisit your prompts every time you switch models.

What changes

The article introduces three new commands you can run from Claude Code.

• /claude-api prompt-audit — scans the prompts, skills, and tool descriptions in your working directory, finds anti-patterns, and removes them
• /claude-api cost-optimize — profiles where your application's tokens go and proposes cost reductions, like unused prompt caching opportunities or work you could batch
• /claude-api hillclimb — splits an evaluation into train and test sets and iteratively searches across models, effort, and prompts to find a configuration with a good cost-performance balance

In a test simulating a migration from Opus 4.8 to Opus 5 on a customer support benchmark, running /claude-api prompt-audit to remove anti-patterns alone dropped costs by an average of 14.6% while accuracy rose by an average of 5.3%. That gain also came bundled with fixing real bugs, like contradictory rules causing the model to withhold refunds it owed, and a retired thinking setting causing the API to reject requests outright.

Calibrating effort matters a lot too. On the FrontierCode Diamond benchmark, Claude Fable 5 scored 11.5% at low effort for $5.35 per task, versus 30.9% at max effort for $19.00 per task. Raising effort costs about 3.5x more but gains about 2.7x the score (+19 points), so it's often worth it. But on Humanity's Last Exam, going from low effort (about 53% accuracy at about $0.30 per question) to max effort (about 61% at about $2.23) barely moved the needle on that last step, the gain fell within the benchmark's run-to-run noise. Knowing where effort actually pays off matters.

Dive Deep

The article walks through a few concrete case studies too.

In one /claude-api hillclimb run on a customer support benchmark, starting from Opus 4.8 at high effort, the search first tried Opus 5 at low effort plus prompt-audit, which held 98.9% train accuracy while dropping cost to 2.6 cents per ticket. Switching to Sonnet 5 at low effort cut cost further to 1 cent per ticket, but accuracy fell to 88.9%. Reading the failing train examples, the search added routing rules and a refund-cap cross-reference to the prompt, bringing Sonnet 5 back to 98.9% accuracy at the same cost. On the held-out test tickets, the final configuration scored 90.5% versus the original setup's 78.6%, at about one-fifth the cost.

/claude-api cost-optimize was also run across four public benchmarks:

• LegalBench: thinking tokens dropped from 102,779 to 8,284, pass rate stayed within noise, and cost fell about 58%
• tau2-bench retail: explicit prompt-cache breakpoint placement alone cut cost about 73% while keeping pass rate flat
• OfficeQA Pro: batch processing and document caching brought cost down from $136.20 to $64.87, about 52% lower
• SWE-bench Verified: dropping effort to medium and constraining output to a few concise sentences cut median steps per task from 29 to 17 and prompt tokens from 75.2M to 33.7M, for about a 55% cost reduction

What struck me across every case was how much of the savings came from unglamorous, stackable tweaks, cache placement, prompt cleanup, and effort tuning, while performance stayed almost untouched.

Wrap-up

• Raise your prompt cache hit rate (keep volatile content out of the prefix, use defer_loading to keep rarely used tools out of the cache, and so on)
• Run /claude-api prompt-audit whenever you switch models, to flush out stale prompt anti-patterns
• Calibrate effort to the task, a stronger model at low effort can beat a weaker model working hard at high effort, and cost less too
• /claude-api cost-optimize and /claude-api hillclimb can automate this whole tuning process for you

If you're building on the Claude API and specifically want to cut cost without giving up performance, this is the kind of thing you'll want to try today!