llm-cost-attr

FinOps cost attribution tool for LLM API usage. Scans your codebase, finds every LLM API call, estimates costs, and recommends optimizations.

Think of it as a cost linter for AI spend.

Install

npx llm-cost-attr scan .

Or install globally:

npm install -g llm-cost-attr
llm-cost-attr scan .

What it does

1. Scans your codebase for LLM API calls across 5 providers
2. Estimates monthly costs based on current model pricing
3. Detects loops, caching patterns, and batch opportunities
4. Recommends optimizations (model downgrades, caching, batching)

Supported Providers

| Provider | Detection Pattern | |----------|------------------| | OpenAI | openai.chat.completions.create(), openai.embeddings.create() | | Anthropic | client.messages.create() | | Google GenAI | model.generateContent(), model.embedContent() | | LangChain | ChatGoogleGenerativeAI, ChatOpenAI, ChatAnthropic, .invoke(), .ainvoke() | | Claude CLI | claude -p subprocess calls |

Works with both TypeScript/JavaScript and Python codebases.

Usage

Scan a repository

llm-cost-attr scan /path/to/repo

Set monthly invocation volume

llm-cost-attr scan /path/to/repo --volume 10000

Output formats

# Terminal (default)
llm-cost-attr scan /path/to/repo

# JSON (for CI pipelines)
llm-cost-attr scan /path/to/repo -f json -o report.json

# HTML dashboard
llm-cost-attr scan /path/to/repo -f html -o report.html

Compare multiple repos

llm-cost-attr compare /repo1 /repo2 /repo3

View pricing

llm-cost-attr pricing

Example Output

  LLM Cost Attribution Report
  ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  Scanned: 193 files | Found: 5 LLM call sites
  Estimated monthly cost: $9.07 @ 1K invocations/month

┌──────────────────────────┬───────────┬──────────────────┬──────────┬──────┬──────────┐
│ File                     │ Provider  │ Model            │ $/call   │ Loop │ $/month  │
├──────────────────────────┼───────────┼──────────────────┼──────────┼──────┼──────────┤
│ src/aggregation.ts:232   │ anthropic │ claude-haiku-4-5 │ $0.0016  │ 5x   │ $8.00    │
│ src/llm-scorer.ts:178    │ anthropic │ claude-haiku-4-5 │ $0.0010  │ 5x   │ $5.00    │
│ src/llm-scorer.ts:163    │ google    │ gemini-2.0-flash │ $0.0001  │ 5x   │ $0.31    │
└──────────────────────────┴───────────┴──────────────────┴──────────┴──────┴──────────┘

  Optimization Opportunities:
  📦  src/aggregation.ts:232
      5 sequential calls in a loop. Consider batching or parallel execution.

How it works

Static analysis — no runtime proxy, no API keys needed. The tool:

1. Walks your codebase (respects .gitignore)
2. Detects SDK imports (OpenAI, Anthropic, Google GenAI, LangChain)
3. Finds API call sites using regex + heuristic analysis
4. Extracts model names, max_tokens, prompt sizes
5. Detects loops (including cross-function call chains)
6. Checks for existing caching (Redis, in-memory)
7. Calculates costs using current provider pricing
8. Generates optimization recommendations

Pricing Data

Built-in pricing for all major models (updated March 2026):

• Anthropic: Opus 4 ($5/$25), Sonnet 4 ($3/$15), Haiku 4.5 ($0.25/$1.25)
• OpenAI: GPT-4o ($2.50/$10), GPT-4o-mini ($0.15/$0.60), o1/o3
• Google: Gemini 2.0 Flash ($0.10/$0.40), Gemini 2.0 Pro ($1.25/$5)

Per 1M tokens (input/output). Run llm-cost-attr pricing for the full list.

License

MIT