$1.00 for a $0.50 Coding Job

01 / The receipt: $0.50 of work, $0.50 of avoidable traffic

Your coding assistant does not need to choose a more expensive model to double your bill. It can do that by rereading files, swallowing terminal noise, explaining too much and retrying without new evidence.

Here is an illustrative receipt for one coding task across multiple requests. These are hypothetical rates, not CheapRelay prices or measured benchmark results: $2 per million input tokens and $10 per million output tokens. All input in this example is uncached. Actual costs depend on the selected model, its current rates and cache treatment.

• ■ Useful task input: 150,000 tokens × $2 per million = $0.30.
• ■ Useful task output: 20,000 tokens × $10 per million = $0.20.
• ■ Lean task total: $0.50.
• ■ Unnecessary repeated context: 100,000 input tokens = $0.20.
• ■ Unfiltered logs: 50,000 input tokens = $0.10.
• ■ Unneeded explanations: 10,000 output tokens = $0.10.
• ■ One avoidable retry cycle: 25,000 input tokens plus 5,000 output tokens = $0.10.
• ■ Bloated task total: $1.00.

The final meter reads 325,000 input tokens and 35,000 output tokens: $0.65 plus $0.35. The extra traffic costs as much as the useful work.

02 / Receipt line: $0.20 for sending the same context again

A common expensive habit is asking the assistant to inspect everything before every small change. Repository trees, unrelated modules and previous discussion can become recurring baggage.

In this receipt, ten requests each carry 10,000 unnecessary input tokens. That is 100,000 tokens, or $0.20 at the illustrative rate. This is an accounting example, not a claim about what every assistant automatically resends.

Ask for a targeted search first, then provide the relevant files and line ranges. Keep a compact task note containing the objective, constraints, decisions and current failure. Do not replace necessary context with a vague summary; remove material that cannot affect the next decision.

03 / Receipt line: $0.10 for terminal output nobody needed

Full test suites and verbose build logs can bury one useful error under thousands of irrelevant lines. Feeding that output back into the assistant makes the clutter billable input.

Our example adds 50,000 such tokens across the task. Instead, run the smallest relevant test and capture the failing assertion, stack trace and enough surrounding output to preserve context. Expand the capture when the first excerpt is insufficient. Truncation should remove noise, not hide evidence.

04 / Receipt line: $0.10 for explanations after every edit

At these hypothetical rates, output costs five times as much per token as input. Asking for a tutorial after each routine patch therefore has a visible price.

The receipt assigns 10,000 unnecessary output tokens to repeated plans, narrated edits and long recaps. Request a short completion report: files changed, tests run, results and unresolved risks. Keep detailed reasoning for decisions you actually need to review.

05 / Receipt line: $0.10 for retrying without a changed premise

A failed attempt is not automatically waste. A second attempt becomes suspect when the assistant repeats the same approach without a new error, hypothesis or constraint.

Here, that loop adds 25,000 input tokens and 5,000 output tokens. Before authorizing another pass, ask what changed. For ambiguous work, set a checkpoint after two unsuccessful attempts: inspect the evidence and approve a revised plan rather than allowing indefinite exploration.

06 / Check cache charges before calling repeated context free

Caching can change this receipt substantially, but a cache hit is not the same as zero cost. Separate uncached input, cache reads, cache writes and output when reviewing usage. Apply the actual rate for each category rather than the simplified rates above.

CheapRelay offers pay-as-you-go access with no monthly fee or subscription, billing per real token, including cache reads and writes. That makes request-level accounting worth doing: a low rate does not eliminate unnecessary volume.

Sign up for CheapRelay to track your own task receipts, then compare similar completed jobs by total spend and acceptable results—not just by the cheapest individual request.