Your AI Bill Is Lying to You. Per-Call Credits Show What Every Prompt Really Costs

A monthly AI invoice tells you one number. It doesn't tell you which prompt, which model, which agent or which runaway loop produced it—so it can't tell you what to change. Per-call metering does. Here's why the unit of AI accounting has to be the call, and how to read your bill once it is.

The Problem With One Big Number

Provider pricing is per token, and the published rates vary by orders of magnitude between models. On OpenAI's price list, gpt-6-luna is $0.10 per million input tokens while gpt-6-astra is $10—a hundredfold gap—and cached input is priced around 90% below standard input. Anthropic's list shows the same shape: Claude Haiku 5.5 at $0.10 input and $0.50 output per million tokens, Claude Sonnet 5.5 at $2 and $10, cache reads at a tenth of base input or less, and a 50% Batch API discount.

With spreads like that, a single monthly total hides everything that matters. One agent on a flagship model with a long system prompt can quietly account for most of the bill while a hundred cheap calls look like the culprit. The invoice is accurate and useless at the same time.

Why Seats Lie Too

Flat subscriptions have the opposite problem: they hide cost by making it constant. A seat costs the same whether the person asked three questions or ran an agent loop all week, so nobody has any reason to use a smaller model, cache a prompt, or cap a loop. Flat pricing feels safe and removes every incentive to be efficient.

What Per-Call Metering Looks Like

The routers that sit in front of many models already expose the right unit. OpenRouter's usage accounting, for instance, returns a per-request usage object with prompt and completion token counts and a cost field for "the total amount charged to your account", included automatically in every response. That's the raw material: a cost attached to every single call, not a total at month end.

MultiverseAI builds the bill on that unit. Every AI call—chat reply, agent step, image, video, speech, search—is recorded as a charge: 1 credit = $0.01, billed at the provider's cost plus 50%, with fractions carried over until they add up to a whole credit. The Credits page shows the last seven and thirty days by feature and by model, and lists recent charges one by one. Calls made with your own provider keys cost zero credits and still show up, so the picture is complete.

How to Read a Per-Call Bill

Once the data is per call, four questions become answerable in a minute each:

1. Which feature costs the most? Sort by feature. If video or image generation dominates, that's a usage decision, not a model decision. If chat dominates, look at the model mix.
2. Which model is doing the cheap work? Sort by model. Summaries, classification and routing should be on the small tier. If a flagship is handling them, move them.
3. Which single calls are outliers? Scan the recent-charges list for anything many times the median. Long contexts and runaway agent loops show up here first.
4. What does one unit of work cost? Divide a workflow's charges by its runs. That number—cost per handled ticket, per generated post, per report—is what you compare against the alternative, and it's the number a finance team actually wants.

Turning the Bill Into Changes

• Route by difficulty. Put a Condition node in front of two AI Agent nodes with different models. Send easy requests to the small one. The price lists above show how much room there is.
• Cache the parts that repeat. System prompts, tool definitions and retrieved context that appear in every call are what cached-input pricing is for.
• Schedule what isn't urgent. Nightly reports and bulk classification can run on a Schedule trigger instead of on demand—and on providers that offer batch pricing, that's where the discount lives.
• Cap every agent. Max iterations, max tokens, max spend per task. Approval gates in front of anything that sends or buys.

Make one change a week and watch the per-call numbers move. If a change doesn't move them, undo it; the bill will tell you within a day.

Honesty Is the Feature

"Your AI bill is lying" is a little unfair—the invoice is telling the truth about the total. It's just answering the wrong question. The right question is "what did that prompt cost?", and the only bill that can answer it is one built from calls. Open the Credits page after your next workflow run and read it the way you'd read a receipt. It changes how you build.

Sources

1. OpenAI, "Pricing" (API) — https://developers.openai.com/api/docs/pricing
2. Anthropic, "Pricing" (Claude Developer Platform) — https://platform.claude.com/docs/en/docs/about-claude/pricing
3. OpenRouter, "Usage Accounting" — https://openrouter.ai/docs/use-cases/usage-accounting