Anatomy of a Claude Bill

Six lines on every invoice

A Claude bill has more lines than 'tokens × price'. Per call you pay for: input tokens, output tokens, cache write tokens (if you used cache_control), cache read tokens (if you got a hit), tool invocation tokens (server tools), and image tokens (computed from dimensions). Each rate differs; planning at the line level prevents surprises.

Find the dominant line

For chat workloads, output tokens often dominate (long answers). For RAG-stuffing workloads, input tokens dominate. For repeated-context workloads with caching, cache reads dominate (cheap) and cache writes are rare. Knowing which line dominates tells you where the next dollar saved comes from.

Track per-feature, not per-bill

The Anthropic Console shows org-level billing. To know which feature drives cost, tag every call with a feature label and aggregate yourself. cwkPippa logs feature: 'chat' | 'heartbeat-weather' | 'council' per call and computes per-feature cost from JSONL — the bill is then explainable to the household budget meeting.

Principle: The bill has shape. Find the dominant line, optimize that, ignore the rest.

Code

Per-call cost calculator·python

# Rates from Anthropic pricing page (USD per million tokens) — verify before deploying.
RATES = {
    "claude-sonnet-4-6": {
        "input": 3.00,
        "output": 15.00,
        "cache_write": 3.75,
        "cache_read": 0.30,
    },
    "claude-haiku-4-5-20251001": {
        "input": 1.00,
        "output": 5.00,
        "cache_write": 1.25,
        "cache_read": 0.10,
    },
}

def cost(usage, model: str) -> float:
    r = RATES[model]
    return (
        (usage.input_tokens / 1_000_000) * r["input"]
        + (usage.output_tokens / 1_000_000) * r["output"]
        + (usage.cache_creation_input_tokens / 1_000_000) * r["cache_write"]
        + (usage.cache_read_input_tokens / 1_000_000) * r["cache_read"]
    )

Feature tagging and aggregation·python

import json, time

def log_call(feature: str, model: str, usage, log_path: str):
    line = {
        "ts": time.time(),
        "feature": feature,
        "model": model,
        "input": usage.input_tokens,
        "output": usage.output_tokens,
        "cache_write": usage.cache_creation_input_tokens,
        "cache_read": usage.cache_read_input_tokens,
        "cost_usd": cost(usage, model),
    }
    with open(log_path, "a") as f:
        f.write(json.dumps(line) + "\n")

Six lines on every invoice

Find the dominant line

Track per-feature, not per-bill

Code

External links

Exercise

Progress

Comments 0