$1 Cap for 1,000 Text Jobs: Check the Receipt

01 / A $1 ceiling before the first request

For a beginner processing 1,000 short texts, the first decision is not which model sounds smarter. It is how much money you can lose while learning. Set a $1 ceiling, reserve $0.20 for mistakes, and plan the work around the remaining $0.80. This is a budget example, not a claim that either model will complete your job for that price.

Use DeepSeek V4 Flash for the first pass and test DeepSeek V4 Pro on difficult records. Do not assume either will meet your quality target without checking. The useful receipt contains accepted results, actual token usage and actual charges—not just the number of requests sent.

02 / Turn 1,000 texts into a token estimate

Choose a narrow task: classify each customer message into one of five categories and return a short reason. Give each record an ID, define the categories, and request only the ID, category and reason. Avoid asking for an essay you will immediately discard.

For planning only, suppose each request consumes 200 input tokens, including instructions, and produces 40 output tokens. Those are illustrative counts, not measured performance figures. Your messages, repeated instructions and response lengths will change them.

• ■ 1,000 records × 200 input tokens = 200,000 input tokens.
• ■ 1,000 records × 40 output tokens = 40,000 output tokens.
• ■ The first-pass planning total is 240,000 tokens.

Input and output can have different prices, so multiplying 240,000 by a single headline rate is not a reliable estimate. Check the current rates and billing categories before sending the batch.

03 / Price the plan without inventing a bargain

Let F-in and F-out represent the current Flash input and output prices per million tokens. The first-pass estimate is 0.20 × F-in + 0.04 × F-out, in the currency of those prices. This formula deliberately leaves prices blank: a useful walkthrough should not turn an unverified rate into a promised receipt.

CheapRelay provides pay-as-you-go access to both models through one key, with no monthly fee or subscription. Billing follows real tokens, including cache reads and writes. If your requests use caching, calculate those categories at their applicable rates rather than treating every input token identically.

Sign up for CheapRelay, check the current model rates and exact API identifiers, and fund only the small amount you are comfortable testing.

04 / Spend the first pilot on 20 representative records

Select 20 records that include ordinary messages, ambiguous wording, long examples and messy formatting. Use the OpenAI-compatible base URL https://api.cheapapirelay.com/v1 in your client, set your key, and submit one record per request initially. Start with low concurrency so mistakes are easy to inspect.

Save each record ID, selected model, returned result, usage data and charge where available. Keep the key outside your source code. Manually review all 20 outputs before scaling: a cheap response with the wrong category is still wasted money.

If those 20 records cost $0.016, multiplying by 50 projects $0.80 for 1,000 comparable records, including the pilot. That is a conditional calculation, not an observed charge. It also leaves no room within the $0.80 allocation for longer records or additional attempts. Use actual pilot spending and leave headroom.

05 / Make Pro earn its extra receipt line

Escalate records using explicit rules: missing fields, invalid categories or ambiguity that your first-pass instructions cannot resolve. Do not rely solely on a model saying it is confident. Test whether Pro actually fixes the examples you care about.

If 100 records receive a second pass with the same illustrative token counts, that adds 20,000 input tokens and 4,000 output tokens. Using Pro prices per million, the added estimate is 0.02 × P-in + 0.004 × P-out. Include the original Flash pass too; escalation does not erase its cost.

06 / Close the receipt before repeating the batch

Reconcile first-pass charges, escalations, retries and cache charges against your account usage. Save completed results so restarting does not resend successful records. Check accumulated spending between small groups of requests; a script-side limit is not a guaranteed provider-side hard cap. Your final measure is total spend divided by accepted records. That is the receipt worth using for the next batch.