Opus Plans, Haiku Types: A $0.318 Receipt

A coding task does not need the same model for every token. Deciding how to change a payment workflow is different from typing the agreed implementation. Give the difficult decisions to Claude Opus 5, then hand a bounded specification to Claude Haiku 4.5. The useful question is not whether that sounds efficient. It is what appears on the receipt.

Here is a worked budget, not a benchmark or a price quote. The rates below are hypothetical accounting inputs, not verified CheapRelay prices. Replace them with the current rates before spending money. Token counts are an example workload, not measured results.

01 / The example task has two separately billed stages

Imagine adding an export endpoint to an existing application. The planning stage examines the relevant code and produces a short implementation contract: files to change, response fields, authorization rules, edge cases and acceptance tests. The typing stage implements that contract.

• ■ Planning: 6,000 input tokens and 2,000 output tokens.
• ■ Implementation: 18,000 input tokens and 12,000 output tokens, aggregated across its calls.
• ■ Illustrative planning-model rates: $15 per million input tokens and $75 per million output tokens.
• ■ Illustrative typing-model rates: $1 per million input tokens and $5 per million output tokens.

The implementation input allowance includes its instructions, relevant source files and the planning handoff. That handoff is not free: it is output when generated and input when another model reads it. This example excludes cache reads and writes, tool charges and retries; it does not assume those items cost zero.

02 / Receipt A: use the planning model for everything

First, price both stages at the hypothetical planning-model rates. Every line uses tokens divided by one million, multiplied by the applicable rate.

• ■ Planning input: 6,000 × $15 / 1,000,000 = $0.090.
• ■ Planning output: 2,000 × $75 / 1,000,000 = $0.150.
• ■ Implementation input: 18,000 × $15 / 1,000,000 = $0.270.
• ■ Implementation output: 12,000 × $75 / 1,000,000 = $0.900.
• ■ Total example token cost: $1.410.

The implementation output is the largest line. In this particular budget, spending less on typing has more room to help than trimming a few sentences from the plan.

03 / Receipt B: keep the plan, change the typing rate

Now keep the planning stage unchanged and price implementation at the hypothetical typing-model rates. Hold token counts constant to isolate the effect of routing; real models may produce different counts.

• ■ Planning input and output: $0.090 + $0.150 = $0.240.
• ■ Implementation input: 18,000 × $1 / 1,000,000 = $0.018.
• ■ Implementation output: 12,000 × $5 / 1,000,000 = $0.060.
• ■ Total example token cost: $0.318.
• ■ Difference from Receipt A: $1.092, approximately 77.4% less.

Across 100 identical example tasks, those totals become $141.00 versus $31.80. That is arithmetic, not a promised saving: it assumes identical workload sizes, no additional review calls and no failed implementations.

04 / A cheap implementation still needs a rejection rule

Make the handoff specific enough to check. Require the implementation stage to stay within named files, preserve public interfaces and pass explicit tests. If it discovers an unresolved design choice, send that question back for planning rather than inviting an architectural guess.

At these assumed rates, another full implementation pass costs $0.078. Another full planning pass costs $0.240. One extra pass of each brings the split workflow to $0.636. Track those additions, plus human review time, instead of reporting only the cheapest successful first attempt.

05 / Keep the CheapRelay receipt tied to actual usage

CheapRelay offers pay-as-you-go access through one key, with no monthly fee or subscription. Its billing covers real tokens, including cache reads and writes. Record those categories separately alongside the model, task identifier and retry count; then reconcile your estimate with billed usage.

Sign up for CheapRelay and try one bounded task with a spending limit. Keep planning and implementation on separate receipt lines. The goal is not the lowest-priced model everywhere, but a lower total cost for work that passes your checks.