$1 Budget: Let Opus Plan and Haiku Type

A tiny AI budget disappears quickly when you send every sentence to the model you chose for the hardest decisions. A cheaper approach is to separate deciding from producing: ask Claude Opus 5 for a compact plan, then give Claude Haiku 4.5 the repetitive writing or implementation work.

Start with a $1 spending envelope. That is a budget you choose, not a claim that a particular project costs $1. The actual bill depends on current model rates, input and output tokens, and any billed cache activity.

01 / Divide your $1 before sending a prompt

For a first experiment, reserve $0.20 for planning, $0.60 for production, and $0.20 for corrections. These are spending limits, not quoted model prices or guaranteed amounts of work. Check the current rates first: if your planned request cannot fit inside its allocation, shorten it or postpone it.

• ■ Planning: one focused request with a short answer.
• ■ Production: a small task, checked before the next request.
• ■ Corrections: money held back for a specific repair.

Keep a simple receipt with the model, task, and actual charge for each request. Unless your tool provides an enforced spending limit, these allocations are manual stopping rules. A note saying “budget: $0.20” in a prompt does not cap your bill.

02 / Give Opus one decision with clear boundaries

Suppose you need a small Python script that turns an exported expense CSV into monthly totals. The expensive mistake would be asking for a complete accounting application. Your first request should define the smallest useful result.

Try this prompt: “Plan a Python script that reads expenses.csv with date, category, and amount columns. It should write monthly totals to totals.csv. Use the standard library. Give me input assumptions, implementation steps, and four acceptance checks. Keep the answer under 300 words. Do not write code.”

The 300-word limit is an instruction, not an exact token limit. If your client supports a maximum output setting, set that too, allowing enough room for a complete answer. Review the plan yourself. Resolve questions such as date format and invalid amounts before paying for implementation.

03 / Hand Haiku the approved plan, not the whole chat

Start a fresh request containing the task, the approved plan, and a tiny sample CSV. There is usually no reason to resend every planning exchange. Repeated context adds input tokens, and longer conversations can make each subsequent turn more expensive.

Your production prompt can say: “Implement the approved plan below as one Python file. Use only the standard library. Return the code and a short command showing how to run it. If a requirement is unclear, identify it instead of adding features.”

This division gives the production model a bounded assignment. It does not guarantee correct code. Run the script on invented sample expenses before using real records, and compare its output with totals you calculate yourself.

04 / Spend the correction reserve on one visible defect

If the script fails, send the relevant error, the smallest input that reproduces it, and the affected code. Ask for a targeted correction. “February totals omit refunds represented as negative amounts” is more useful than “make this better.”

Return to the planning model when the design itself needs reconsideration. A missing import usually needs a small repair; conflicting requirements may need another planning pass. Check your remaining budget before either request. Stop when the acceptance checks pass.

05 / Read the receipt before repeating the workflow

CheapRelay provides access to both models through one key, with OpenAI-compatible routes under /v1. Billing is per real token, including cache reads and writes, with no monthly fee or subscription. Sign up for CheapRelay when you are ready to run your first budgeted experiment.

Compare the total charge with your $1 envelope, including retries. This split saves money only when cheaper production outweighs the added planning and handoff costs. For a two-sentence task, planning may add unnecessary expense. For work with several implementation steps, a short approved plan gives you something concrete to test before you spend again.