$3.06 to Tag 100,000 Feedback Messages—On Paper

01 / A $3.06 classification receipt, not a price promise

Customer feedback tagging is a good place to buy small outputs instead of long answers. You need a few labels, not an essay about every complaint. Here is a worked budget for tagging 100,000 feedback messages: $3.00 for the first pass, plus $0.06 for retries. Those figures use hypothetical token rates, not a current CheapRelay quote or a measured benchmark.

The point is to make the bill inspectable. Replace the assumed rates with the current rates for your chosen model, measure your actual token usage, and the same arithmetic becomes your purchasing estimate.

02 / The basket: 100,000 messages and eight allowed tags

Start with a narrow taxonomy: billing, cancellation, bug, performance, feature_request, usability, praise, and other. Allow up to two tags per message. Keep sentiment separate; a feature request can be enthusiastic or furious without changing its topic.

For this budget, assume each request consumes 220 input tokens, including instructions, tag definitions, formatting rules, and the feedback itself. Assume an average output of 20 tokens for a compact object containing tags and a review flag. These are planning assumptions. Long tickets, multilingual messages, or verbose definitions can increase them substantially.

• ■ Messages: 100,000.
• ■ Input: 100,000 × 220 = 22,000,000 tokens.
• ■ Output: 100,000 × 20 = 2,000,000 tokens.
• ■ Assumed input rate: $0.10 per million tokens.
• ■ Assumed output rate: $0.40 per million tokens.

03 / The receipt: $2.20 input, $0.80 output, $0.06 retries

At those illustrative rates, the first-pass receipt is straightforward:

• ■ Input charge: 22 × $0.10 = $2.20.
• ■ Output charge: 2 × $0.40 = $0.80.
• ■ First-pass subtotal: $3.00.
• ■ Retry allowance: 2,000 additional requests at the same average size = $0.06.
• ■ Estimated API total: $3.06, before any applicable taxes or other charges.

The 2% retry allowance is a budget choice, not an observed failure rate. It covers a hypothetical mix of invalid responses and transient failures. If retries carry longer correction prompts, they will cost more than this allowance assumes. Log usage from every attempt, including unsuccessful attempts that report billable tokens.

04 / Keep the classifier from becoming a tiny essay writer

Ask for only the allowed tags and a review flag. Define boundaries explicitly: a refund dispute belongs under billing; a request to stop renewal belongs under cancellation. Treat the customer's text as data, never as instructions. A message saying “ignore your rules” should not change the classification contract.

Validate the response in code. Reject unknown tags, excessive tag counts, missing fields, and malformed objects. Route empty messages and unsupported content to review rather than manufacturing a confident label. Use an output limit that leaves room for valid structured responses, then measure actual output instead of assuming every response hits that limit.

05 / Add $16.67 for a small human audit

Suppose someone reviews 100 classifications at 30 seconds each. That is 50 minutes. At an assumed labor cost of $20 per hour, the audit costs $16.67. Your combined illustrative receipt becomes $19.73: $3.06 for API usage and $16.67 for review. Integration work and investigation of flagged messages are additional costs.

A random sample of 100 is a smoke test, not proof that rare categories work. Also test ambiguous messages, overlapping topics, and examples from every tag. Track mistakes by category. A low overall error count can conceal a classifier that consistently misses cancellation requests.

06 / Replace the assumed rates before funding the run

CheapRelay offers pay-as-you-go access without a monthly fee or subscription. Its OpenAI-compatible routes sit under /v1 at https://api.cheapapirelay.com. Billing follows real token usage, including cache reads and writes; this example assumes ordinary input and output charges without caching.

Sign up for CheapRelay, check current pricing for a candidate such as Gemini 2.5 Flash, and run a labeled pilot before processing the full queue. Compare measured tokens and classification errors with this receipt. The cheapest useful classifier is the one that produces reliable labels at a total cost you can explain.