Pay for 2 Examples Before You Pay for 10

01 / Two examples are a starting point, not a magic number

Every example in a prompt is extra input you may pay to send again. Before collecting ten polished demonstrations, try two: one ordinary case and one case that exposes a likely mistake. Then check whether they beat clear instructions alone.

Few-shot prompting means showing a model a small number of sample inputs and desired outputs. It can help with classification, extraction and consistent formatting. But examples are not automatically worth their token cost. On a tiny budget, the useful question is: does another example fix enough errors to justify carrying it on future requests?

02 / Start with a zero-example support classifier

Suppose you need to label incoming messages as billing, technical or other. Your first prompt should define those labels, handle overlap and specify the output. For example: “Return exactly one label: billing, technical or other. Payment and invoice questions are billing. Product failures are technical. If both appear, choose the main requested action.”

That is a zero-shot prompt: instructions without demonstrations. Test it before adding anything. If the task already works reliably enough for your needs, the cheapest useful number of examples may be zero.

If it struggles, add this ordinary demonstration: “Input: Please send my invoice again. Output: billing.” Then add a boundary case: “Input: I paid yesterday, but the app crashes on launch. Please fix the crash. Output: technical.” The second example teaches a decision boundary rather than repeating the easy case.

03 / Price the repeated tokens before expanding the prompt

Assume each demonstration uses 80 tokens, including its labels and separators. This is a planning assumption; measure your actual prompt with a tokenizer or API usage data.

• ■ One example adds approximately 80 input tokens per request.
• ■ Two examples add approximately 160 input tokens per request.
• ■ Four examples add approximately 320 input tokens per request.
• ■ Ten examples add approximately 800 input tokens per request.

Moving from two examples to ten adds 640 tokens per request. Across 1,000 requests, that is 640,000 additional input tokens. At a hypothetical uncached input price of $1 per million tokens, the difference is $0.64. That rate is arithmetic practice, not a CheapRelay price quote. Substitute the current rate for your chosen model.

This calculation covers only the added input. Your full bill also depends on the remaining prompt, output and any separately billed token categories. Longer examples can cost much more than these short demonstrations.

04 / Run an 80-request test before buying more context

Build a small test set of 20 realistic messages with labels you have checked yourself. Include straightforward requests, overlapping categories and vague wording. Keep these messages separate from your prompt demonstrations.

Run the same 20 messages with zero, one, two and four examples. That makes 80 requests. Hold the model, instructions, output limit and generation settings constant so you are mostly measuring the examples, not unrelated changes.

• ■ Count correct labels out of 20.
• ■ Count outputs that break the required format.
• ■ Record billed tokens and cost for each variant.
• ■ Inspect which individual messages changed labels.

On this test, one additional correct answer changes the score by five percentage points. That is a coarse signal, not proof of a stable improvement. If two and four examples tie, prefer two provisionally. If four fixes an important error, check that benefit on fresh messages before keeping it.

05 / Replace weak examples before adding a fifth

When a prompt fails, first ask whether the instructions are ambiguous or the demonstrations are repetitive. Replace a redundant easy example with a relevant boundary case. Do not paste the entire test set into the prompt: you would lose your independent check and increase recurring input costs.

CheapRelay offers pay-as-you-go API access with no monthly fee or subscription, which suits small experiments. Billing follows real token usage, including cache reads and writes; do not assume repeated examples are free. Sign up for CheapRelay when you are ready to measure a small test against current model pricing.

The practical stopping rule is simple: keep the smallest example set that meets your task's needs on fresh inputs. Start with zero, try two, and make every additional example earn its place.