Track token usage and cost per feature and customer

Approach

Every response includes usage. Record input, output and cache tokens with the model, feature and customer for each call, compute cost with your price table, and alert on unusual spend.

Solution

import anthropic
from dataclasses import dataclass

client = anthropic.Anthropic()

# $ per million tokens; keep this table in config and update it when prices change
PRICES = {"claude-sonnet-5-5": {"in": 3.00, "out": 15.00, "cache_read": 0.30, "cache_write": 3.75}}


@dataclass
class UsageRecord:
    feature: str
    customer_id: str
    model: str
    input_tokens: int
    output_tokens: int
    cache_read: int
    cache_write: int

    @property
    def cost(self) -> float:
        p = PRICES[self.model]
        return (self.input_tokens * p["in"] + self.output_tokens * p["out"]
                + self.cache_read * p["cache_read"] + self.cache_write * p["cache_write"]) / 1e6


def tracked_create(feature: str, customer_id: str, **params):
    response = client.messages.create(**params)
    u = response.usage
    record = UsageRecord(feature, customer_id, params["model"], u.input_tokens, u.output_tokens,
                         u.cache_read_input_tokens or 0, u.cache_creation_input_tokens or 0)
    save_usage(record)                          # write to your metrics or database
    return response

Pitfalls

• Hard-coding prices in many places; keep one table.
• Ignoring cache tokens, which are billed at different rates.
• Only looking at the monthly invoice, after the damage is done.

Variations

• Enforce per-customer daily budgets before calling the model.
• Export usage as metrics (Prometheus, OpenTelemetry) and build a dashboard by feature.

Further reading: Pricing