The Receipt Behind a 12-Call Tool Agent

01 / One task, twelve model calls, one receipt

A tool-heavy agent is not buying one answer. It is buying repeated decisions over an expanding conversation. The useful comparison is therefore cost per completed task, not simply price per million tokens.

Consider a deliberately simplified workflow: twelve model calls, starting with 2,000 input tokens. Each subsequent call carries another 1,000 tokens of accumulated instructions, tool results, and conversation. Each call produces 300 output tokens. Total usage is 90,000 input tokens and 3,600 output tokens.

Every rate below is an illustrative budgeting assumption, not a verified provider or CheapRelay quote. These are arithmetic scenarios, not performance benchmarks. Actual billed usage can differ, including separately metered reasoning or caching categories.

02 / The all-routine-model receipt: $0.0558

Assume a routine model costs $0.50 per million input tokens and $3.00 per million output tokens. With no caching, its receipt is:

• ■ Input: 90,000 ÷ 1,000,000 × $0.50 = $0.0450.
• ■ Output: 3,600 ÷ 1,000,000 × $3.00 = $0.0108.
• ■ Model subtotal: $0.0558 per workflow.
• ■ At 10,000 identical workflows: $558.00.

The input total matters. Twelve calls do not mean twelve copies of the initial 2,000-token prompt. Here, inputs grow from 2,000 to 13,000 tokens. Adding another tool result can affect several later calls, not just the next one.

03 / Two escalations raise the receipt to $0.1255

Now assume an escalation model costs $3.00 per million input tokens and $15.00 per million output tokens. Route the first ten calls to the routine model and the last two to the escalation model. This is a fixed comparison scenario, not a claim that the final two calls always need more capability.

• ■ Routine input: 65,000 tokens × $0.50 per million = $0.0325.
• ■ Routine output: 3,000 tokens × $3.00 per million = $0.0090.
• ■ Escalation input: 25,000 tokens × $3.00 per million = $0.0750.
• ■ Escalation output: 600 tokens × $15.00 per million = $0.0090.
• ■ Mixed-model subtotal: $0.1255 per workflow, or $1,255.00 for 10,000.

Using the escalation model for all twelve calls would cost $0.3240 per workflow, or $3,240.00 for 10,000. The mixed route saves $1,985.00 against that scenario, but costs $697.00 more than the all-routine route. Whether that premium is worthwhile depends on your measured completion rate.

04 / Tool fees and retries belong on separate lines

Suppose each workflow also makes five paid searches at an assumed $0.002 each and uses one sandbox execution costing an assumed $0.005. Those external tool charges add $0.0150. They are separate hypothetical expenses, not stated CheapRelay charges.

The mixed-route total becomes $0.1405 per workflow, or $1,405.00 for 10,000. If 8% of those workflows each require exactly one complete rerun, with identical usage and fees, the total becomes $1,517.40. That assumption excludes further failures and partial retries.

Track accepted completions alongside spending. A lower token bill is not a better result if it produces more invalid tool arguments, unnecessary searches, or tasks that require manual repair.

05 / Choose with your own task-level receipts

For a candidate such as Gemini 2.5 Flash, test your actual tool schemas and representative tasks rather than assuming a low token price guarantees low task cost. Record model calls, input and output usage, tool charges, retries, and whether the result passes your acceptance checks.

Keep uncached input, cache reads, cache writes, and output on separate receipt lines. This example assumes no caching. Do not apply a cache discount to the entire conversation: use the usage categories and rates actually billed.

CheapRelay provides a pay-as-you-go way to run that comparison through one key, with OpenAI-compatible routes and support for existing OpenAI SDK workflows. Billing follows real token usage, including cache reads and writes, without a monthly fee or subscription. Sign up for CheapRelay to compare a small batch using current rates before committing a larger workload.

Choose the least expensive route that reliably passes your checks. Keep escalation explicit, tool output bounded, and the full receipt attached to each completed task.