$0.65 instead of $2.27: split planning from typing

01 / A $0.65 receipt, not a promised API price

Putting every step through your most expensive model is easy. It also means paying planning prices for repetitive implementation. A cheaper arrangement is to let Claude Opus 5 define the work, let Claude Haiku 4.5 produce the routine output, and reserve one final expensive pass for checking the result.

Here is a worked example: $0.65 for the split workflow against $2.27 for the same token volumes handled entirely by the planning model. Those are illustrative costs, not a quote, benchmark, or claim about current CheapRelay rates. The point is to make the arithmetic visible before you build the routing.

02 / The rate card behind this example

For this exercise, assume the following prices per million tokens. These are deliberately hypothetical budgeting inputs; replace them with the current rates for your chosen models before estimating a real job.

• ■ Planning model: $10 per million input tokens and $50 per million output tokens.
• ■ Typing model: $1 per million input tokens and $5 per million output tokens.
• ■ No cache reads or writes in this example, and no retries in the initial receipt.

The calculation is simple: input tokens multiplied by the input rate, plus output tokens multiplied by the output rate, with both token counts divided by one million. Input and output need separate lines because a long generated patch can cost more than the instructions that requested it.

03 / One plan, six implementation tasks, one review

Imagine a small application change divided into six bounded tasks. The planner reads the requirements and relevant code, then produces acceptance criteria, file boundaries, interface decisions, and test expectations. Each implementation call receives the plan excerpt and code needed for its task, rather than the whole planning conversation.

• ■ Planning: 20,000 input tokens and 2,000 output tokens on the planning model. Cost: $0.20 + $0.10 = $0.30.
• ■ Implementation: six calls, each using 10,000 input tokens and 4,000 output tokens. Combined typing-model cost: $0.06 + $0.12 = $0.18.
• ■ Final review: 12,000 input tokens and 1,000 output tokens on the planning model. Cost: $0.12 + $0.05 = $0.17.
• ■ Total: 92,000 input tokens, 27,000 output tokens, and a $0.65 bill under these assumptions.

The review allowance is not unlimited. If the combined patch and necessary context exceed 12,000 input tokens, increase that line. A receipt stops being useful when it quietly excludes the material the reviewer must actually read.

04 / The $1.62 saving comes from implementation

Run those same assumed token volumes entirely through the planning model and the bill becomes $0.92 for input plus $1.35 for output: $2.27 total. Moving only the six implementation calls reduces their cost from $1.80 to $0.18. Planning and review remain unchanged at $0.47.

That is $1.62 less, or about 71.4%. It is a controlled arithmetic comparison, not evidence that both workflows will produce identical token counts or equally good code. Different models can need different context, generate longer answers, or require additional attempts.

05 / Give retries their own receipt line

One repeated implementation task adds $0.01 for input and $0.02 for output at the assumed typing rates. Two such retries add $0.06, taking the split bill to $0.71 before any additional review.

Use tests and explicit acceptance criteria to decide when a task is finished. Escalate ambiguous requirements, architectural changes, and repeated failures to the planner. Do not keep purchasing cheap attempts after the problem has stopped being routine.

06 / Measure the split on your actual bill

CheapRelay provides pay-as-you-go access through one key, with no monthly fee or subscription. Billing follows real token usage, including cache reads and writes, so record those categories separately when caching enters your workflow.

Sign up for CheapRelay and try this split on one bounded change. Log tokens, retries, review corrections, and the final charge. Keep the arrangement only if the completed, checked result costs less—not merely because the first drafting call looked cheap.