← Retour aux guides

Cut AI Output Costs: A Practical Prompt-Tuning Guide

Optimize the response, not just the model price

A cheaper API rate helps, but unnecessary output can quietly erase the savings. If your application needs a classification, returning three explanatory paragraphs is wasted spending. Prompt tuning means changing instructions and examples; it does not require training or fine-tuning model weights.

Start by defining the smallest response your application can actually use. For a support router, that might be one category and a confidence score. For a code assistant, it might be a patch and two implementation notes rather than a tutorial.

1. Measure a baseline before editing prompts

Run 100 representative requests and record input tokens, output tokens, latency, and task success. Use API usage fields rather than estimating from character counts. Include difficult requests: a shorter response is not a saving if it causes retries or manual correction.

Suppose your current request averages 1,500 input tokens and 900 output tokens. At illustrative rates of $1 per million input tokens and $5 per million output tokens, each request costs $0.006. These are example rates, not quoted 59API prices. At 10,000 requests, the baseline bill is $60.

2. Replace vague brevity requests with a response contract

“Be concise” leaves the model guessing. Specify the deliverable, length, and exclusions instead. For a support summary, try: “Return exactly three bullets: problem, likely cause, next action. Each bullet must contain no more than 15 words. No introduction, repeated question, or closing offer. If a fact is missing, write ‘unknown’ rather than inventing it.”

For classification, request only the allowed label. For extraction, specify the exact fields and exclude commentary. Where your model and endpoint support structured outputs, use a schema to improve consistency; validate the returned data either way.

• Replace: “Explain your recommendation in detail.”
• With: “Give one recommendation and one sentence explaining the main trade-off.”
• Replace: “Review this code.”
• With: “List up to three actionable defects. Include file, line, and fix. Omit style preferences.”

3. Add a token ceiling without breaking answers

Set the endpoint’s supported output limit alongside your instructions. Claude Messages uses max_tokens; OpenAI-compatible endpoints may use max_tokens or max_completion_tokens, depending on the model. Check the selected endpoint rather than assuming one parameter works everywhere.

For a brief summary, test a 200-token ceiling against your previous 1,000-token allowance. A ceiling limits generation; it does not guarantee a complete answer. Monitor stop reasons and reject truncated responses. On reasoning models, the generation budget may also include reasoning tokens, so visible response length alone can understate billed usage.

4. Calculate the savings, including prompt overhead

Assume the revised instructions add 50 input tokens while output falls from 900 to 150 tokens. Using the same illustrative rates, the new cost is (1,550 × $1 + 150 × $5) ÷ 1,000,000 = $0.0023 per request.

At 10,000 requests, spending drops from $60 to $23: a $37 saving, or approximately 61.7%. That calculation includes the extra instructions. Long examples can eat into savings, so keep only examples that measurably improve accuracy.

5. Test quality and retries, then choose your route

Compare both prompts on the same 100-request dataset. Track valid-format rate, correctness, average cost, and retries. If the shorter prompt saves 60% per call but repeatedly produces incomplete answers, optimize cost per successful task instead.

59API offers low-cost, pay-as-you-go access to native, official-quality Claude and GPT models without downgrades. Its compatibility with Claude Code, Codex, and OpenAI SDKs makes it practical to test prompt changes without rebuilding your application. Configure https://api.59api.com as the API base URL, following the endpoint requirements for your client.

Compare current per-model rates with your measured token usage; inexpensive models are especially worth testing for classification and extraction. Check referral rebate terms separately rather than counting rebates as guaranteed savings. If you are ready to benchmark your shorter prompts, consider signing up for 59API and starting with a small, capped test budget.

Tarifs · Docs