LLM API Cost Reduction: Caching, Routing, Token Budgets
Breaks your LLM bill down per call site, then ranks fixes like output trimming, caching, batch, and routing by savings and risk.
Before & after
Same prompt, same agent — the only difference is the skill.
That's a significant budget. There are definitely ways to optimize that spend. First, I'd want to understand where the costs are coming from—which models are you using, what's your usage pattern, how many tokens per request? Are there specific workflows that are particularly expensive? You might benefit from model selection strategy, prompt caching, or batch processing. It really depends on your architecture. Tell me more about what you're building, and I can suggest some approaches to bring that down.
LLM COST OPTIMIZATION ANALYSIS Current spend: 5,000 USD/month | Target: 2,000 USD/month (60% reduction) COST BREAKDOWN • Claude Opus (50%): 2,500 USD/month → 500M input tokens, 50M output tokens → Avg cost per request: 2.15 USD (high complexity tasks) • Claude Sonnet (30%): 1,500 USD/month → 800M input tokens, 100M output tokens → Avg cost per request: 0.42 USD (standard tasks) • GPT-4 (20%): 1,000 USD/month → Legacy integrations, planned retirement OPTIMIZATION #1: Model Downgrade (20% savings) • Analysis: 40% of Opus requests are for routine summarization (complexity score <3) • Action: Route summary tasks to Sonnet (4x cheaper, similar quality) • Implementation: Adjust routing logic in middleware • Estimated savings: 1,000 USD/month OPTIMIZATION #2: Prompt Caching (25% savings) • Finding: System prompts are 8KB each, repeated in 70% of requests • Action: Implement prompt cache with Anthropic API v1.11+ • Setup: Cache policy TLL = 5 min, Markdown formatting (5KB cache per prompt) • Estimated savings: 1,250 USD/month OPTIMIZATION #3: Batch Processing (15% savings) • Finding: Reports and exports run nightly, wait < 12 hours for results • Action: Use batch API (50% discount) instead of streaming • Setup: Queue generation jobs, submit at 2am UTC, retrieve at 10am • Estimated savings: 750 USD/month OPTIMIZATION #4: Max Tokens Enforcement (10% savings) • Finding: No max_tokens limit, some outputs run 8KB (wasting budget) • Action: Set max_tokens=1000 for summary tasks, 2000 for content • Implementation: Default in all API calls • Estimated savings: 500 USD/month IMPLEMENTATION ROADMAP • Week 1: Deploy model routing (Sonnet for summaries) • Week 2: Enable prompt caching • Week 3: Batch API setup (reports, exports) • Week 4: Max tokens enforcement EXPECTED OUTCOME: 2,100-2,200 USD/month (56% reduction)
About this skill
Your OpenAI or Claude bill doubled and nobody can say which feature caused it. Turning on caching changed nothing, and someone wants to move everything to the cheap model. This skill works out where each dollar goes, then ranks the fixes by savings against quality risk and gives you the exact code and config changes.
What it does
- Breaks spend down per call site into calls, uncached input, cached input, and output tokens, so you work on the term that actually dominates.
- Restructures prompts so the cache hits: orders tools, system prompt, and static examples ahead of the cache point, catches prefix-breakers like timestamps and unstable tool JSON, and runs the write-versus-read break-even.
- Cuts output tokens with enum labels, structured outputs, and reasoning-budget limits, which often saves more than prompt shortening does.
- Routes work to cheaper models only behind an eval gate, including the cascade math that shows when escalation makes routing cost more.
- Controls agent and chat loops, where resending the full history makes cost grow roughly with the square of turn count.
- Specifies per-feature and per-tenant cost logging and quotas, so the next spike can be traced and a single abuser cannot drain the budget.
What you get
A cost breakdown table, a ranked list of levers with estimated monthly savings, quality risk, and effort for each, the concrete prompt, cache, and routing changes, and the metrics to watch after rollout.
Best for
Engineers and tech leads running LLM features in production or pre-launch, especially support triage, extraction, chat, and agent workloads where the API bill has become a line item someone asks about.
SKILL.md preview
---
name: llm-api-cost-reduction
description: Cuts LLM API spend by breaking cost down per call site (calls x uncached input, cached input, and output tokens), then applying fixes in order of savings per unit of risk. The fixes are output trimming, prompt caching, batch, eval-gated model routing, and agent-loop context control, plus per-feature cost logging. Use when the user says "my OpenAI bill tripled", "our Claude API costs are out of control", "which model should I use for this", "set up prompt caching", "my cache hit rate is zero", "what will this AI feature cost at launch", "our agent burns through tokens", or "how many tokens is this request".
---
… (sign up to view the full skill)Featured in
More ai & agents skills
View all AI & Agents skills →New Skill Scaffolder
Create a new Claude Code skill correctly: right mechanism, valid name and folder, a triggering description, and a smoke test.
Skill Library Lint and Overlap Report
Lint every skill in a repo at once: frontmatter, name-folder match, weak descriptions, overlapping triggers, broken refs, plus a CI script.
Self-Verifying Agent Loop: Plan, Check, Retry, Close
Turns a goal into a task plan with machine-run checks, executes with retry caps, escalates, and refuses to close until verified.
Skill Test Runner (Trigger and A/B)
Test a skill in fresh sessions: ready-to-run prompts, commands, and rubric for trigger recall, false triggers, and blind A/B.