The Model Ladder: A $10.24 AI Receipt

01 / Put the expensive model behind a quality gate

A model ladder is a spending rule: send a task to a cheaper model first, check the result, and escalate only when it fails a defined requirement. The goal is not the lowest price per token. It is the lowest total cost per usable result.

Here is a receipt for 10,000 product-description extraction jobs. Every job turns a supplied paragraph into structured fields. All rates, token counts, and failure rates below are illustrative planning assumptions, not CheapRelay prices, measured benchmarks, or promises about any model.

02 / The hypothetical rate card behind the receipt

• ■ First rung: $0.20 per million input tokens and $0.80 per million output tokens.
• ■ Escalation rung: $2.00 per million input tokens and $8.00 per million output tokens.
• ■ Initial request: 1,200 input tokens and 300 output tokens.
• ■ Escalated request: 1,800 input tokens and 400 output tokens, allowing space for the failed answer and validation feedback.
• ■ Planning assumption: 8% of initial answers fail validation, so 800 jobs escalate.

Use the actual rates and measured token counts from your account before budgeting production traffic. A larger retry prompt is intentional here: escalation rarely costs exactly the same as the original request.

03 / Receipt: 10,000 jobs, with 800 escalations

• ■ First-rung input: 12 million tokens × $0.20 = $2.40.
• ■ First-rung output: 3 million tokens × $0.80 = $2.40.
• ■ Escalation input: 1.44 million tokens × $2.00 = $2.88.
• ■ Escalation output: 0.32 million tokens × $8.00 = $2.56.
• ■ Total model spend: $10.24.

Sending every job directly to the escalation rung, using the original 1,200-input and 300-output payload, would cost $24.00 for input plus $24.00 for output: $48.00 total. Under these assumptions, the ladder reduces model spend by $37.76, or 78.7%.

That is a bill comparison, not proof of equivalent quality. The receipt excludes hosting, engineering time, human review, taxes, and any additional retries. If escalated answers still fail, their resolution belongs on the bill too.

04 / Make failure observable before you route traffic

For this extraction task, require valid JSON, all mandatory fields, allowed category values, and prices that match the supplied text. Reject invented product attributes. A missing source value should remain missing according to your schema, not become a confident guess.

Run deterministic checks first. They add no model-token charges, although they still consume compute and maintenance time. Some requirements need human sampling or a separate evaluation process; valid JSON alone cannot establish factual correctness.

Do not ask the first model whether its answer is good enough and treat that response as your quality gate. Define acceptance outside the generation step. Record the failure reason so the next rung receives specific repair instructions rather than a vague request to try harder.

05 / The escalation rate can erase the saving

In this example, every escalation costs $0.0068. The initial pass costs $4.80 regardless of how many answers fail. Your budget equation is therefore $4.80 plus the number of escalations multiplied by $0.0068.

At a 30% escalation rate, model spend becomes $25.20. At 70%, it becomes $52.40, exceeding the $48.00 direct route. The mathematical break-even point is approximately 63.5%, before accounting for extra latency or unresolved failures. Measure this rate by task type, not just across your whole application.

06 / Use one relay without hiding the token bill

CheapRelay gives you one key for multiple model providers, with pay-as-you-go billing and no monthly fee or subscription. That makes it a practical place to implement routing without maintaining separate provider credentials for each rung.

Billing follows real token usage, including cache reads and writes. This receipt assumes no caching; do not quietly count cached tokens as free. Sign up for CheapRelay, test a small representative batch, and replace every assumption in this receipt with observed costs and acceptance results before scaling.