$1.20 for 60 Accepted Answers Is Not the Best Deal

01 / The $1.20 receipt loses to the $1.55 receipt

A smaller API bill does not necessarily buy cheaper results. Suppose one route spends $1.20 and produces 60 accepted answers. Another spends $1.55 and produces 90. The first costs $0.0200 per accepted answer. The second costs about $0.0172: roughly 14% less, despite the larger bill.

That is the useful comparison: total evaluation spend divided by accepted answers. Not price per million tokens alone. Not the cost of one impressive response.

Every price, token count and acceptance result below is hypothetical arithmetic, not a published benchmark or a CheapRelay price quote. The purpose is to show a receipt you can reproduce with your own usage.

02 / Set the acceptance rule before spending anything

Start with 100 representative tasks from your actual workload. Use the same tasks, available context, tools and acceptance rubric for both routes. Define an accepted answer before seeing the outputs.

For a structured extraction job, acceptance might require valid JSON, all required fields and exact agreement with verified source values. For a code change, it might require passing specified tests without prohibited edits. Fluent writing alone should not count as success unless that is genuinely the job.

Also set a retry policy. In this example, each route may retry a failed task once using the same correction prompt. Count each task as accepted at most once. Multiple successful attempts on one task do not create multiple accepted answers.

03 / Receipt A: $0.90 initially, then $0.30 in retries

For the budget route, assume the following first-pass token charges across all 100 tasks. Rates are dollars per million tokens. Each line is a separate, non-overlapping billing bucket.

• ■ Uncached input: 600,000 tokens at $0.50 = $0.30.
• ■ Cache reads: 400,000 tokens at $0.10 = $0.04.
• ■ Cache writes: 100,000 tokens at $0.60 = $0.06.
• ■ Output: 250,000 tokens at $2.00 = $0.50.
• ■ First-pass subtotal: $0.90; 50 tasks accepted.
• ■ Retries on failed tasks: $0.30 in additional token charges; 10 more tasks accepted.
• ■ Final receipt: $1.20; 60 unique tasks accepted.

The result is $1.20 divided by 60, or $0.0200 per accepted answer. Dividing by all 100 tasks would produce $0.0120 per task attempted, but that hides the 40 tasks still unresolved.

04 / Receipt B: $1.55 with no retries needed

For the second route, assume the same first-pass token volumes but higher hypothetical rates. It accepts 90 tasks immediately; its permitted retries produce no additional acceptances or token charges because the workflow stops rather than retrying those failures.

• ■ Uncached input: 600,000 tokens at $1.00 = $0.60.
• ■ Cache reads: 400,000 tokens at $0.20 = $0.08.
• ■ Cache writes: 100,000 tokens at $1.20 = $0.12.
• ■ Output: 250,000 tokens at $3.00 = $0.75.
• ■ Final receipt: $1.55; 90 unique tasks accepted.

Its accepted-answer cost is $1.55 divided by 90, approximately $0.0172. The higher-priced token schedule wins this particular evaluation. That does not establish a universal model ranking.

05 / Keep unresolved work beside the winning number

Publish acceptance coverage alongside cost: 60% for the first route, 90% for the second. A route that solves only easy tasks can look economical while leaving expensive cleanup behind. For production decisions, also measure the complete workflow, including fallback calls, tool charges and human review.

Use actual billed usage rather than estimating everything from prompt length. Cache eligibility, output length and retries can change the receipt. Repeat the evaluation and report variability; 100 tasks are a starting sample, not proof of lasting superiority.

06 / Build the receipt from your own relay usage

CheapRelay offers pay-as-you-go access through one API key, with no monthly fee or subscription. Billing covers real tokens, including cache reads and writes, so keep those categories separate when calculating accepted-answer cost.

Sign up for CheapRelay, set a small evaluation budget, and compare complete receipts on your own tasks. Buy the lowest cost per usable result that meets your acceptance and coverage requirements—not merely the lowest advertised token rate.