Four Bill Lines for a Heavy-Workload Team

A heavy-workload team cannot manage its API budget from a single total labeled “tokens.” Two calls with similar prompt lengths can cost different amounts because one reuses cached context while the other creates it. Output length adds another variable. For Quanta Briefings, this walkthrough follows the four billing categories that make those differences understandable: input, output, cache write and cache read.

01 · Step 1: Identify the model before interpreting the charge

Start with the model recorded for the call and the applicable rates. Do not carry a rate assumption from one model into another. A workload using Claude Sonnet 5.5 and Gemini 2.5 Flash needs separate cost calculations, even when both receive the same task.

Quanta meters calls on real tokens, with cache writes and cache reads priced separately. Its pay-as-you-go balance covers multiple models, which is convenient for sustained workloads, but a shared balance does not mean a shared rate.

02 · Step 2: Separate ordinary input from cached context

Input tokens represent material sent into the model: instructions, conversation history, retrieved documents, tool definitions and other request content. The important accounting question is whether a displayed input total includes cached tokens or reports ordinary input separately.

Do not automatically add every token counter together. Some usage formats expose cached tokens as a subset of an input total. Others provide separate categories. Establish what each field means before multiplying it by a rate; otherwise, the same prompt tokens can be counted twice.

03 · Step 3: Read output as its own spending line

Output tokens are billed separately from input. A short question can produce a long answer, and an automated coding or analysis task can generate substantial output across repeated calls.

Use the billable output count, not the visible answer’s apparent length. Depending on the model and usage format, billable generation can include tokens that are not presented as ordinary answer text. When investigating a costly call, compare its reported output usage with its task, output limits and completion behavior.

04 · Step 4: Treat a cache write as context preparation

A cache write records eligible context for potential reuse. It is not a cache hit, and it is not automatically free. Read its token quantity and applicable write rate separately from ordinary input.

For a team repeatedly supplying a stable instruction block or reference document, that write may support cheaper later calls. However, the economics depend on subsequent reuse, cache lifetime and the model’s caching rules. A block written once and never reused still incurs its applicable write charge.

05 · Step 5: Verify cache reads rather than assuming reuse

A cache read indicates that eligible context was reused. The read quantity should be charged at the applicable cache-read rate, not automatically at the ordinary input rate.

Repeated content alone does not guarantee a hit. Changes to a reusable prefix, cache expiration or model-specific eligibility requirements can prevent reuse. Check the reported read count. If an expected hit is absent, investigate request construction before forecasting savings across the entire workload.

06 · Step 6: Reconstruct the call using consistent units

Once the categories are understood, calculate each line independently:

• › Ordinary input: billable ordinary input tokens multiplied by the input rate.
• › Output: billable output tokens multiplied by the output rate.
• › Cache write: cache-write tokens multiplied by the applicable write rate.
• › Cache read: cache-read tokens multiplied by the applicable read rate.

If rates are quoted per million tokens, divide each token quantity by one million before multiplying. Confirm whether cache-write pricing replaces ordinary input pricing for those tokens or represents an additional charge. Then sum the applicable charges without duplicating tokens or fees.

07 · Step 7: Turn the calculation into a rollout budget

Track cold calls, cache-hit calls and long-output calls separately. Budget from their expected mix rather than assuming every request will resemble the cheapest successful run. Include retries and multi-call tasks when estimating the cost of completed work.

Quanta is a practical fit for teams that want one pay-as-you-go balance while comparing workloads across models. Open a Quanta account and start with a bounded workload, reconcile its four billing lines, then expand only when observed usage supports the budget.