How to track LLM costs per customer

Updated August 20, 2026.

"$14,300 this month" answers nothing. Which customer drove it? Which feature? Is your biggest account profitable at its current plan? For AI products — where inference is a real cost of goods sold and usage-based pricing is winning — per-customer cost attribution has become table stakes. Industry surveys put it bluntly: fewer than half of teams can attribute AI spend to a customer at all.

Provider dashboards can't do it. OpenAI's and Anthropic's usage APIs report org-level daily dollars — useful for reconciling an invoice, useless for knowing that acme-corp consumed 45% of yesterday's spend.

The recipe#

Attribution means capturing four things at every LLM call site, at the moment the response arrives:

1. The provider — openai, anthropic, gemini, … It namespaces the price lookup.
2. The response-reported model — not the alias you requested. Responses echo dated snapshots (gpt-4o-2024-08-06), and snapshots price differently.
3. The usage object, as the provider sent it — token counts, including cache and reasoning details.
4. Your attribution fields — at minimum customer and feature; teams running agents add a run/session id, multi-tenant products add end-user.

Cost is then tokens × a price table, computed server-side, and stored frozen at ingest — never recomputed when prices change later.

The traps (each one is a real, documented bug class)#

• Inclusive vs exclusive input tokens. Anthropic's input_tokens excludes cached tokens; OpenAI's prompt_tokens includes them. Naive cross-provider math double-bills cached tokens — errors up to 2x have shipped in popular open-source trackers.
• Streaming loses usage. OpenAI streams only report usage if you set stream_options: {"include_usage": true}, on a final extra chunk. Anthropic sends cumulative totals in the last event — summing deltas overcounts. Capture at end-of-stream, from the provider's own numbers.
• Unknown models must never be $0. A new model or a custom fine-tune that isn't in your price table yet should land as unpriced, visibly — a silent zero corrupts every average built on top of it.
• Stale price tables. Providers reprice quietly. Several corrections on our own pricing pages this month came from the community price catalog lagging an official change — a maintained, synced table is part of the job.

Build vs buy#

The capture side is genuinely small — one call after each LLM response. The grind is everything after: maintaining prices for thousands of model variants, normalizing five usage dialects, handling batch/cache/tier modifiers, and keeping a queryable store. That second half is what Marginal sells:

marginal.track({
  provider: "openai",
  model: response.model,        // response-reported, snapshot included
  usage: response.usage,        // as-is; dialects normalized server-side
  fields: { customer: "acme-corp", feature: "support-bot" },
});

One buffered, fire-and-forget call per response. Cost is computed at ingest against a daily-synced catalog (with your custom prices for anything private), unknown models land flagged as unpriced, and the dashboard slices spend by any field you registered — customer, feature, model, whatever your margins depend on.

Start tracking, or read the docs for the full event shape.