$0.144 for One Reasoning Run—If the Rates Fit

01 / A 14.4-cent receipt, not a live price quote

A reasoning-heavy prompt can produce a short answer and still generate a substantial bill. Comparing Gemini 3.1 Pro and Claude Opus 5 therefore starts with the usage receipt, not the length of the answer on screen.

Here is a concrete accounting example: 12,000 input tokens and 10,000 billable output tokens. At illustrative rates of $2 per million input tokens and $12 per million output tokens, that request costs $0.144. Those rates are arithmetic assumptions, not verified prices for either model. Check the current model rates before treating any calculation here as a purchasing estimate.

02 / The $0.144 request, line by line

Assume a difficult code-review prompt containing source files, constraints, and a request to identify a subtle failure. For this worksheet, suppose the usage report records 8,000 reasoning tokens and 2,000 visible answer tokens, both billed at the assumed output rate.

• ■ Input: 12,000 tokens × $2 / 1,000,000 = $0.024.
• ■ Reasoning: 8,000 tokens × $12 / 1,000,000 = $0.096.
• ■ Visible answer: 2,000 tokens × $12 / 1,000,000 = $0.024.
• ■ Total: $0.024 + $0.096 + $0.024 = $0.144.
• ■ Batch of 100 identical requests: $14.40.

Reasoning accounts for two-thirds of this hypothetical bill. That is not a benchmark result or a claim about either model's typical behavior. It shows why counting only the final answer can badly understate expenditure.

Usage fields vary by endpoint. If reasoning tokens are already included in the reported output total, do not add them again. This receipt splits one output total into two components; it does not charge the same tokens twice.

03 / The model comparison starts with two actual receipts

Run the same task through Gemini 3.1 Pro and Claude Opus 5, then record each model's actual input, billable output, cache usage, and applicable rates. Identical prompts do not guarantee identical token counts or reasoning effort.

For an uncached request, the calculation is straightforward: input tokens multiplied by the input rate, plus billable output tokens multiplied by the output rate, with both rates divided by one million.

At the worksheet's fixed token counts, every additional $1 per million input tokens adds $0.012. Every additional $1 per million output tokens adds $0.010. A $10 difference in output pricing therefore changes this receipt by $0.10, even before differences in generated token volume enter the comparison.

04 / One retry turns $0.144 into $0.288

A cheap attempt is not necessarily a cheap completed task. Two identical attempts at the example rates cost $0.288; three cost $0.432. Across 100 tasks, that is $28.80 or $43.20 instead of $14.40.

Measure cost per accepted result: total spending divided by results that pass your checks. For code review, those checks might require a reproducible failure and a valid fix. Do not assume either model needs fewer retries without testing your own workload.

05 / Cache lines need their own arithmetic

If 9,000 of the 12,000 input tokens qualify as cache reads, separate those from the remaining 3,000 input tokens. Price each category at its applicable rate and include any separately billed cache writes. Never charge all 12,000 tokens at the ordinary input rate and then add cache reads on top.

CheapRelay bills per real token, including cache reads and writes, with no monthly fee or subscription. That makes a small, pay-as-you-go comparison practical without adding a recurring platform charge to this worksheet.

06 / Buy accepted answers, not impressive token totals

Set a small spending cap, preserve the usage records, and compare both quality and total task cost. Where supported, constrain reasoning effort and answer length separately, then check whether the savings survive your acceptance tests.

Sign up for CheapRelay to run that comparison through one API key. Use current rates and actual usage—not this illustrative rate card—to decide which model earns the next request.