A $140 API Bill Cut to $40 by Sending Less

01 / The $100 charge hiding inside unnecessary context

A customer asks whether an order has shipped. Your application sends the order status, but also the customer's profile, saved addresses and entire support conversation. Most of that information does not help answer the question. It still enters the prompt, and billable input still costs money.

Here is an illustrative monthly receipt: 10,000 requests, each carrying 12,000 input tokens and producing 500 output tokens. Assume $1 per million input tokens and $4 per million output tokens. These are hypothetical planning rates, not a CheapRelay price quote or a measured workload.

• ■ Input: 10,000 requests × 12,000 tokens = 120 million tokens, costing $120.
• ■ Output: 10,000 requests × 500 tokens = 5 million tokens, costing $20.
• ■ Total: $140.

Now remove 10,000 unnecessary input tokens from each request. Keep the useful context and the same output allowance. The modeled bill becomes $40. That is $100 saved, or 71.4% of the original total, without assuming cheaper output or a different model.

02 / Separate the 2,000 useful tokens from the 10,000 extras

For this example, allocate the remaining input budget explicitly. These are illustrative token allocations, not claims about how many tokens a particular customer record contains.

• ■ Task instructions: 400 tokens.
• ■ Relevant business rules: 400 tokens.
• ■ Response structure: 200 tokens.
• ■ Selected order facts and necessary context: 1,000 tokens.
• ■ Total retained input: 2,000 tokens.

The removed allocation consists of 5,000 tokens of unrelated conversation, 3,000 tokens of profile material and 2,000 tokens of saved addresses and delivery notes. None is needed for this particular shipment-status answer. Another task might legitimately require an address; minimization means selecting by purpose, not banning fields universally.

The revised receipt is straightforward: 20 million input tokens cost $20, and the unchanged output costs $20. Total: $40. Input spending falls 83.3%; the whole bill falls 71.4%. Those percentages differ because the output charge remains.

03 / Build the prompt from an allowlist, not a customer dump

Start with the question and decide which facts can change the answer. For shipment status, that might mean the latest fulfillment event, its timestamp and the expected delivery window. Fetch those fields rather than serializing the customer object and trying to clean it afterward.

Keep account identifiers in application logic unless the model needs them. If the application can attach the answer to the correct order itself, the prompt may not need an email address, full name or internal customer identifier.

Redaction and omission solve different problems. Replacing an email address with a placeholder removes that address from the text, but leaves the surrounding record. If the entire record is irrelevant, omit it. Also check the request envelope: removing data from the main message does not help if a tool result or attachment includes it again.

04 / Count cache charges without calling caching privacy

The $140-to-$40 receipt assumes ordinary uncached input billing. Do not apply it unchanged to a workload with cache hits. CheapRelay bills real tokens, including cache reads and writes, so an actual receipt must separate those categories and use their applicable rates.

A cache can change the cost of repeated context; it does not make unnecessary user data necessary. Removing a block from future requests also does not establish that previous copies have disappeared from logs, caches or downstream systems. Review those separately.

05 / Verify the smaller request before claiming the saving

Compare the original and reduced prompts on representative tasks. Check answer correctness, missing context, clarification requests and output length. If trimming causes retries, include those extra calls in the receipt. Use recorded token usage rather than estimating production savings from character counts.

CheapRelay's pay-as-you-go billing, with no monthly fee or subscription, makes request-level accounting useful. Sign up for CheapRelay when you are ready to test the smaller payload. Keep the first target simple: send only the user data the task needs, then verify both the answer and the bill.