Budgeting 100,000 Calls by Input and Output

A team planning 100,000 API calls needs more than an average cost per request. Reading a large document and returning a short classification creates a different bill from reading a brief instruction and generating a long report. At sustained volume, that difference shapes model selection, cache strategy and spending limits. Work through the budget in this order.

01 · Step 1: Split the workload by what each call consumes

Start with two illustrative workloads. These are planning examples, not measured customer results:

• › Document review: 20,000 input tokens and 1,000 output tokens per call.
• › Report generation: 2,000 input tokens and 8,000 output tokens per call.

Across 100,000 calls, document review consumes 2 billion input tokens and 100 million output tokens. Report generation consumes 200 million input tokens and 800 million output tokens. The first workload processes more total tokens, but that alone does not establish which costs more. Input and output have separate rates, and cached input introduces additional billing categories.

02 · Step 2: Price the token categories separately

For each candidate model, calculate cost from its applicable billing categories: uncached input, cache writes, cache reads and output. Multiply each category’s token count by its rate, adjusting for the rate’s unit, then add the charges. Use the provider’s billing definitions so that cached tokens are not accidentally counted again as ordinary input.

Without caching, the examples produce a useful comparison. Document review costs more when the output token rate is less than roughly 2.57 times the input token rate. Report generation costs more when that ratio is higher. This threshold comes directly from the illustrative token volumes; it is not a claim about any model’s current pricing.

Build this calculation for each model under consideration. A lower input rate can matter greatly for document review, while a lower output rate can dominate the report budget.

03 · Step 3: Measure which input can actually be reused

Input-heavy work offers opportunities for caching when calls share a qualifying prompt prefix. Repeated instructions, reference material or repository context may be reusable. Unique documents still need processing, so a large prompt does not automatically imply large cache savings.

Record cache writes and reads separately. A write incurs its applicable charge; subsequent reads may change the economics if the cached content remains available and requests meet the model’s caching requirements. Estimate savings from observed reuse, including cache expiry and prompt changes, rather than assuming every repeated passage receives a cache hit.

04 · Step 4: Give output-heavy jobs a useful stopping point

For report generation, inspect whether the requested output length serves the task. Specify required sections, remove redundant summaries and set an output limit that accommodates a complete answer. A lower limit only helps if it reduces unnecessary generation without causing incomplete results and retries.

Measure the full job, including validation failures, continuations and regeneration. A short answer that repeatedly needs repair can consume more tokens than a complete first attempt. Likewise, compare models on acceptable completed reports, not just their advertised output rates.

05 · Step 5: Run a representative batch before choosing a model

Test realistic document lengths, report sizes and repeated context. If Claude Sonnet 5.5 and DeepSeek V4 Pro are candidates, evaluate both against the same acceptance criteria and calculate cost per accepted result. Include difficult cases that are likely to trigger retries.

Quanta fits naturally into this evaluation: one pay-as-you-go balance supports multiple models, with every call metered on real tokens and cache writes and reads priced separately. It works with Claude Code, Codex, Gemini CLI and any OpenAI SDK. Open a Quanta account to measure representative batches before committing the full workload.

06 · Step 6: Set budgets around the workload’s main cost driver

For input-heavy queues, monitor context growth, cache reuse and repeated document submission. For output-heavy queues, monitor generated length, completion quality and retry volume. Track both average job cost and expensive outliers, then expand traffic in stages.

Keep separate forecasts for these workloads as volume grows. Their costs respond to different changes, so each needs its own spending assumptions and operational limits.