LLM cost per user: how to track AI cost by customer

What makes up the AI cost of one user?

Model tokens are only part of it. Count every paid step a user's request sets off. Prices checked November 10, 2026.

• Input and output tokens, priced separately. Output usually costs more: Claude Sonnet 5.5 is $2 in and $10 out per million tokens.
• Cached input. A cache hit on Sonnet 5.5 costs $0.10 per million tokens, but writing to the cache costs more than plain input.
• Built-in tools. Web search on the Claude API costs $10 per 1,000 searches, plus the tokens the results add.
• Paid APIs your agent calls, such as search, scraping or enrichment.
• Retries and failed runs, which use tokens without producing anything the user sees.

How do I attribute each call to a customer?

Pick one ID per paying account and send it on every call. Tools differ in how they capture it.

Observability tools estimate cost from token counts and a price table. Langfuse notes it cannot infer cost for reasoning models without token counts, and that price changes apply only to new data. A billing ledger records what was actually charged.

What do light, typical and heavy users cost?

Assume each request sends 5,000 input tokens and gets 500 output tokens back from Claude Sonnet 5.5. Some requests also use web search on the Claude API, at $10 per 1,000 searches. The plan price is $20 a month.

Averages hide the heavy user. If 1 user in 10 is heavy and the rest are typical, the average cost is $8.19 and looks fine, while one account loses money every month.

How do I keep margin on heavy users?

Combine a cheaper path for routine work with limits customers can see.

• Route simple requests to a cheaper model. On Claude Haiku 5.5, at $0.10 and $0.50 per million tokens, the heavy user above costs $1.50 in model fees instead of $30.
• Cache repeated context, and send work that can wait to a batch API. Anthropic and OpenAI both price batch work at half the standard rate.
• Set a monthly budget per user, and show it in the product before it runs out.
• Include an allowance in the plan and bill usage above it.
• Charge paid tool calls at cost plus a markup instead of folding them into the plan price.

• Per-seat vs usage-based pricing
• Cheapest way to call AI models

How do I track cost per customer with Locus Pro?

On a Locus Pro Enterprise workspace, each paid call can name the customer with X-Locus-End-User. Every response returns the charge and an API call ID, and calls made with a secret key also return the base cost. Model calls through the OpenAI-compatible endpoint report input, output and reasoning tokens when the provider sends them.

The ledger can be filtered by customer, provider and endpoint, and customer analytics lists each customer with their spend. That gives you cost per customer from what was charged, not from an estimate.

You can cap each customer with a policy: a budget per hour, day, week or month, a maximum charge per call, and allowed or blocked models. Locus Pro sees only calls made through it. Calls you send straight to a provider need an observability tool or your own logs.

• Locus Pro for platforms
• Control AI agent tool spending

Which way of tracking AI cost should I use?

Use the charged amounts when you bill; use estimates only to explore.

• You need traces, prompts and quality data along with cost: an observability tool such as Langfuse or Helicone.
• You bill customers for the usage: take the amounts from your billing records, not from estimates.
• Your agents call paid third-party APIs for customers: Locus Pro records and bills each call per customer.
• You call providers directly at low volume: log the provider usage fields with your user ID and keep your own price table.