$3.45 for 95 accepted answers beats cheap tokens

01 / The receipt that changes the ranking

A lower token price does not guarantee a lower bill for useful work. If a cheap model needs more retries and still leaves more tasks unfinished, its apparent advantage can disappear. The useful comparison is cost per accepted answer: total billed inference cost divided by the number of tasks that actually pass your acceptance checks.

Consider two hypothetical routes processing the same 100 tasks. Route A costs $0.02 per attempt. Route B costs $0.03. Judged by attempt price alone, A looks like the obvious bargain. After retries and acceptance checks, however, A costs $0.0457 per accepted answer, while B costs $0.0363.

These are illustrative assumptions, not provider prices, measured benchmarks or claims about any particular model.

02 / Start with an explicit $0.02 attempt

For this example, every attempt uses exactly 8,000 input tokens and 4,000 output tokens. That includes retries. There are no cache reads, cache writes, tools or other charges in this simplified receipt.

• ■ Route A input: 8,000 tokens at an assumed $1 per million = $0.008.
• ■ Route A output: 4,000 tokens at an assumed $3 per million = $0.012.
• ■ Route A total per attempt: $0.020.
• ■ Route B input: 8,000 tokens at an assumed $1.50 per million = $0.012.
• ■ Route B output: 4,000 tokens at an assumed $4.50 per million = $0.018.
• ■ Route B total per attempt: $0.030.

Real attempts rarely have identical token counts. A retry may carry a longer conversation, and one model may produce longer answers. Use actual billed usage in a real comparison; fixed lengths here simply keep the arithmetic inspectable.

03 / Route A: $3.20 buys 70 accepted answers

Give each task one initial attempt and, if it fails, one retry. Route A passes 40 tasks on its first attempt. The remaining 60 receive retries, and 30 of those pass.

• ■ Initial attempts: 100 × $0.02 = $2.00.
• ■ Retry attempts: 60 × $0.02 = $1.20.
• ■ Total API spend: $3.20.
• ■ Accepted tasks: 40 + 30 = 70.
• ■ Unresolved tasks: 30.
• ■ Cost per accepted answer: $3.20 ÷ 70 = approximately $0.0457.

The rejected attempts remain on the receipt. Dividing only the cost of successful calls by successful answers would hide the money spent reaching them.

04 / Route B: $3.45 buys 95 accepted answers

Route B passes 85 tasks immediately. Its remaining 15 tasks receive one retry each, and 10 pass.

• ■ Initial attempts: 100 × $0.03 = $3.00.
• ■ Retry attempts: 15 × $0.03 = $0.45.
• ■ Total API spend: $3.45.
• ■ Accepted tasks: 85 + 10 = 95.
• ■ Unresolved tasks: 5.
• ■ Cost per accepted answer: $3.45 ÷ 95 = approximately $0.0363.

B costs $0.25 more overall but delivers 25 more accepted answers. Its cost per accepted answer is about 21% lower, calculated using the unrounded results. Its attempt price is 50% higher. Both statements are true; they measure different things.

05 / Define acceptance before opening the bill

Acceptance must mean the same thing for both routes. For a coding task, require the same tests to pass. For extraction, require the same fields and correctness checks. For writing, use a predefined rubric and review outputs without revealing their source where practical.

Use the same task set, retry limit and stopping rule. Count each task at most once as accepted. Report acceptance rate alongside unit cost: a route that handles only easy tasks can look economical while leaving an expensive backlog. Record latency and human review separately, or add their measured costs if you are calculating total operating cost.

06 / Keep the token bill auditable at CheapRelay

CheapRelay provides pay-as-you-go API access with no monthly fee or subscription. Billing follows real tokens, including cache reads and writes. When comparing routes, preserve those charges rather than treating cached input as free. Record the task identifier, route, attempt number, billed cost and acceptance result for every call.

Sign up for CheapRelay to run a small, budget-capped comparison on your own workload. Rank the resulting receipts by cost per accepted answer, then check completion rate against your requirements. Buy accepted work, not merely inexpensive attempts.