$2.83 Instead of $13.50: A Two-Step Model Ladder

01 / The $2.83 receipt starts with explicit assumptions

Sending every request to your most expensive model is easy. It also means paying premium rates for tasks that might need nothing more than valid JSON and three extracted fields. A model ladder reverses that default: start cheap, check the result, and climb only when the result fails a defined quality gate.

Here is a worked budget for 1,000 tasks. Every price below is hypothetical planning arithmetic, not a CheapRelay quote or a claim about any model’s performance. Replace these rates with current prices before spending money.

• ■ Each attempt uses 2,000 input tokens and 500 output tokens.
• ■ The cheap tier costs $0.20 per million input tokens and $0.80 per million output tokens.
• ■ The higher tier costs $3 per million input tokens and $15 per million output tokens.
• ■ All 1,000 tasks start cheap; an assumed 150 fail validation and receive one higher-tier attempt.
• ■ There are no cache operations, tool charges, or additional retries in this example.

02 / Every failed cheap attempt stays on the receipt

A cheap attempt costs (2,000 × $0.20 + 500 × $0.80) / 1,000,000 = $0.0008. A higher-tier attempt costs (2,000 × $3 + 500 × $15) / 1,000,000 = $0.0135.

• ■ Cheap pass: 1,000 attempts × $0.0008 = $0.80.
• ■ Escalation pass: 150 attempts × $0.0135 = $2.025.
• ■ Total ladder token cost: $2.825, displayed as $2.83.
• ■ Higher tier for every task: 1,000 × $0.0135 = $13.50.
• ■ Difference before rounding: $10.675, or approximately 79.1% less token spend.

The $0.80 includes the 150 cheap attempts that failed. Those tokens are not refunded just because another model handles the task afterward. The assumed 15% escalation rate is not a benchmark, and this receipt does not establish equivalent final quality.

03 / A failed check, not a vague feeling, triggers escalation

Define acceptance before choosing the ladder. For invoice extraction, require parseable JSON, the exact field set, a supported currency, and totals that reconcile within a stated tolerance. For code changes, use relevant tests and explicit requirements. A response that merely sounds polished should not pass automatically.

Separate deterministic checks from judgment calls. A schema validator can reject a missing field cheaply. It cannot reliably establish that a nuanced explanation is correct. Use representative human-reviewed examples to test that part of the gate, including cases where a convincing answer is wrong.

For a trial, you could evaluate DeepSeek V4 Flash as the first rung and Claude Sonnet 5.5 as the escalation rung. Treat those as candidates, not a proven ranking. Measure acceptance, final correctness, and actual token use on your own workload before adopting the route.

04 / The second attempt needs its own token budget

This receipt assumes the higher tier receives exactly 2,000 input tokens and produces exactly 500 output tokens. Forwarding the failed answer, validator messages, or conversation history increases that bill. Record those tokens rather than quietly treating escalation as a fresh request of identical size.

Cap the ladder at one escalation initially. If the second attempt fails, send the task to a review queue or return a clear failure. Unlimited retries turn a budget control into an unpredictable loop. Also measure latency: escalated tasks wait for two attempts and validation, not one.

05 / Use one relay key, but keep a task-level ledger

CheapRelay provides pay-as-you-go access through one key, with OpenAI-compatible routes under /v1 at api.cheapapirelay.com. There is no monthly fee or subscription. Billing follows real token usage, including cache reads and writes, so reconcile your ledger against actual usage rather than assuming every request matches this example.

Log the task identifier, rung, acceptance result, input tokens, output tokens, cache usage, and billed cost. Sign up for CheapRelay to test a small, capped batch, then compare the ladder with a higher-tier-only baseline on both cost and final quality. Keep the ladder only if the measured trade-off works.