$0.192 for a coding session, not $0.480

01 / A $0.192 receipt with the assumptions attached

Do not pay the premium-model rate for every file search, test stub, and explanation. Start your pair-programming session with a cheaper model. Escalate a narrow problem only when progress stops. The useful question is not whether the expensive model is better. It is how much of your session actually needs it.

Here is a worked budget, not a measured benchmark or a current price quote. Assume a cheaper tier costs $1 per million input tokens and $4 per million output tokens. Assume a premium tier costs $5 per million input tokens and $20 per million output tokens. These deliberately illustrative rates are not advertised prices for any named model. Check live rates before spending.

• ■ Cheaper-model work: 40,000 input tokens × $1 per million = $0.040.
• ■ Cheaper-model responses: 8,000 output tokens × $4 per million = $0.032.
• ■ Premium escalation: 12,000 input tokens × $5 per million = $0.060.
• ■ Premium responses: 3,000 output tokens × $20 per million = $0.060.
• ■ Total model bill: $0.192.

For comparison, sending the same aggregate 52,000 input tokens and 11,000 output tokens entirely to the premium tier would cost $0.480. The difference is $0.288, or 60%. That is arithmetic under fixed assumptions, not evidence that every real coding task will save 60%.

02 / Spend the first $0.072 on a bounded coding task

Suppose you are fixing duplicate webhook processing. Give the cheaper model the handler, the relevant database constraint, one failing test, and the expected behavior. Ask it to reproduce the failure, propose the smallest patch, and add a regression test. Do not begin by uploading the entire repository.

A cheaper option such as DeepSeek V4 Flash can be your first pairing partner; reserve a premium option such as Claude Opus 5.5 for escalation after checking their current rates. The workflow matters more than either name. Keep acceptance criteria concrete: processing the same event twice must create one record, preserve the response contract, and pass the relevant tests.

Let the cheaper model inspect nearby code and attempt a repair. Run the tests yourself or through your coding tool. A confident explanation is not a passing test, and a passing narrow test does not replace review.

03 / Make two failed repairs the escalation trigger

Set the stopping rule before the session becomes a loop. For this example, escalate after two attempted fixes fail the same regression test, or when the proposed change crosses a security or data-integrity boundary you cannot confidently review. Two attempts is a budget rule, not a universal optimum.

Send the premium model a compact handoff: the failing assertion, relevant code, attempted patches, observed results, and the constraint that must remain true. In this receipt, that handoff and its follow-up consume 12,000 input and 3,000 output tokens. Those tokens are included in the $0.120 escalation charge; they are not free overhead.

Ask for a diagnosis and minimal patch rather than a fresh implementation. Once the blocker is resolved, return routine test expansion and cleanup to the cheaper model. Escalation should be a detour, not a permanent switch.

04 / Keep the receipt honest when context grows

The $0.192 example assumes no cache reads or writes and excludes tool execution, taxes, and developer time. Real sessions may use different token counts on different models. Repeated context uploads, long answers, and failed retries can erase the apparent saving. Track actual input, output, cache charges, and accepted fixes rather than comparing headline token rates alone.

CheapRelay offers pay-as-you-go access through one key, with no monthly fee or subscription. Its OpenAI-compatible API routes sit under /v1 at api.cheapapirelay.com, and billing follows real token usage, including cache reads and writes. That makes a session-level ledger more useful than a monthly allowance mentality.

Sign up for CheapRelay, verify the current rates, and try this workflow on one bounded bug. Keep the cheaper attempts, escalation handoff, and final bill in your notes. Repeat only if the lower API spend survives the cost of your review time.