Cheap First, Escalate Later: A Coding Cost Receipt

01 / One coding task, two spending lanes

Using your strongest model for every coding turn is like paying a specialist to rename variables. Sometimes necessary; usually an expensive default. A cheaper-first workflow gives routine work to a lower-cost model and buys stronger reasoning only when the task produces evidence that it needs it.

Consider a small job: add pagination to an existing endpoint, update its tests, and fix an edge case involving empty results. Start with a lower-cost model for repository inspection, a proposed patch, and test iteration. Escalate a tightly packaged problem if that loop stalls.

The receipt below is an illustrative budget, not a CheapRelay price quote or a measured benchmark. Token volumes and rates are hypothetical assumptions so you can reproduce the arithmetic. Replace them with the applicable rates and actual usage before making a purchasing decision.

02 / The receipt: $0.355 instead of $1.075

Assume the lower-cost lane charges $1 per million input tokens and $5 per million output tokens. Assume the escalation lane charges $5 and $25, respectively. This example includes no separately billed cache reads or writes.

• ■ Routine work, lower-cost input: 120,000 tokens × $1 / 1,000,000 = $0.120.
• ■ Routine work, lower-cost output: 12,000 tokens × $5 / 1,000,000 = $0.060.
• ■ Escalation, stronger-model input: 20,000 tokens × $5 / 1,000,000 = $0.100.
• ■ Escalation, stronger-model output: 3,000 tokens × $25 / 1,000,000 = $0.075.
• ■ Total cheaper-first bill: $0.355, or about $0.36.
• ■ Same assumed token volumes entirely at stronger-model rates: $0.700 input + $0.375 output = $1.075.

The difference is $0.720, approximately 67% of that comparison bill. This is a same-volume accounting comparison, not proof that two models would require identical conversations. A stronger model might finish sooner. A cheaper model might need more retries. Your real receipt decides whether the strategy worked.

03 / Give the cheaper model a bounded assignment

Do not begin with “fix pagination everywhere.” Specify the endpoint, expected response shape, relevant files, and test command. Ask for the smallest patch that preserves existing behavior outside the requested change. Keep business requirements explicit so neither model has to infer them.

A candidate such as Claude Haiku 4.5 can occupy the cheaper lane if its current price and performance fit your repository. The role matters more than the name: use a model that can make useful progress on your actual tasks, then check the resulting diff and tests.

Before another attempt, inspect what changed. Did a failing test disappear for the right reason? Did the patch introduce unrelated edits? A passing command is evidence, but it does not replace reviewing authorization, data handling, or compatibility where those concerns apply.

04 / Escalate after two failed repairs, not ten vague turns

Use a practical stopping rule: after two repair attempts on the same failure without meaningful progress, pause. Escalate sooner when the model cannot explain the failure, repeatedly reverses its own changes, or proposes bypassing a safeguard. Two attempts is a workflow limit, not a universal optimum.

Send the stronger model a compact handoff: the intended behavior, relevant code, exact error, current diff, and approaches already rejected. Ask for a diagnosis before another patch. The 20,000-token escalation input in the receipt must include that handoff; it is not free context.

Once the blocker is resolved, return routine cleanup to the cheaper lane. Escalation should purchase a specific answer, not silently become the permanent default.

05 / Keep cache charges and developer minutes on the receipt

CheapRelay offers pay-as-you-go access through one API key, without a monthly fee or subscription. That makes it convenient to switch lanes without maintaining separate integrations. Sign up for CheapRelay and run a small, capped trial on representative coding tasks.

Record input, output, cache-read, and cache-write tokens separately because CheapRelay bills real usage, including cache reads and writes. Apply each applicable rate rather than treating cached context as free.

Also record retries and your review time. Saving $0.72 is not a win if it creates twenty minutes of avoidable debugging. Keep the cheaper-first default only when the combined receipt—API charges plus human effort—supports it.