$3.60 for 10,000 Documents: A Worked Receipt

01 / The $3.60 receipt starts with explicit assumptions

Ten thousand documents do not automatically require a large AI budget. For straightforward classification and field extraction, the cheapest sensible approach is usually local preprocessing, a low-cost model, short structured answers, and selective review. Paying a powerful model to read everything twice is not a cost strategy.

Here is a worked budget, not a live price quote. Assume each document sends 1,200 input tokens, including instructions, and receives 200 output tokens. Use illustrative rates of $0.20 per million input tokens and $0.60 per million output tokens. These are arithmetic assumptions, not advertised CheapRelay rates or a claim about any particular model.

• ■ Documents: 10,000.
• ■ Input: 10,000 × 1,200 = 12 million tokens.
• ■ Input charge: 12 × $0.20 = $2.40.
• ■ Output: 10,000 × 200 = 2 million tokens.
• ■ Output charge: 2 × $0.60 = $1.20.
• ■ Base API receipt: $3.60, or $0.00036 per document.

That number is useful because every assumption is visible. Replace the two rates and average token counts with your actual measurements before approving a budget.

02 / Retries and escalation turn $3.60 into $4.03

A production estimate needs room for failed requests and difficult documents. Suppose 2% of documents require one full retry at the same token counts and rates. That adds 200 requests and $0.072.

Now suppose 100 documents also need a second pass through a stronger model. For this example only, price that pass at $2 per million input tokens and $6 per million output tokens. Keep the same 1,200-token input and 200-token output allowance.

• ■ Base processing: $3.60.
• ■ Full retries on 200 documents: $0.072.
• ■ Escalation input: 120,000 tokens, costing $0.24.
• ■ Escalation output: 20,000 tokens, costing $0.12.
• ■ Total before rounding: $4.032.
• ■ Budget receipt, rounded: $4.03.

The escalation charge is additional: those 100 documents already received their first pass. The retry and escalation percentages are planning assumptions, not measured failure rates. Replace them after testing a representative sample.

03 / Short answers save $4.80 in this example

For field extraction, request only the required fields and a compact status value. Do not request an explanation unless a reviewer will actually use it.

At the illustrative output rate above, increasing the average answer from 200 to 1,000 tokens adds eight million output tokens across the batch. That costs another $4.80, taking the base receipt from $3.60 to $8.40 without changing the input. Shorter answers must still preserve required information; truncating valid results is not a saving.

Repeated instructions also count. A 300-token instruction block sent with every document contributes three million input tokens, or $0.60 at our assumed rate. Keep it concise. Do not assume repeated text is free: cache reads and writes can carry charges, and eligibility depends on the route and model.

04 / Skip unnecessary AI calls before shopping for rates

Extract embedded text locally before considering image-based processing. Use deterministic rules for exact matches, known templates, and required-field checks. Deduplicate identical files when the same result can safely be reused. If these steps eliminate 2,000 calls with the assumed average token counts, they remove $0.72 from the base receipt before retry or escalation adjustments.

Test the remaining workflow on representative documents, including awkward scans and missing fields. Compare usable results, not just response prices. DeepSeek V4 Flash is one candidate to evaluate; check its current rate and suitability rather than assuming it matches this example.

05 / Keep the API receipt separate from the full job cost

The $4.03 example excludes OCR, storage, network charges, taxes, engineering time, and human review. Those expenses can outweigh token charges. Record them separately instead of presenting a small API bill as the entire project cost.

CheapRelay offers pay-as-you-go access with no monthly fee or subscription, billing real tokens, including cache reads and writes. Sign up for CheapRelay to test a small sample, then use your measured token usage and current rates to price the full 10,000-document run. The cheapest defensible workflow is the one that meets your quality requirements with the fewest paid tokens and repeat passes.