← Volver a las guías

Cheaper AI Prompts: A Token-Saving API Walkthrough

AI bills often grow because prompts are repetitive and outputs are allowed to expand without a clear limit. The fix is rarely a more complicated prompt. It is usually a smaller input, a narrower task, and an explicit response contract. This walkthrough shows how to reduce token usage while keeping answers useful, using 59API as a low-cost relay for official Claude and GPT models.

Start by measuring the baseline

Before changing a prompt, record the model, input size, output size, latency, and whether the answer passed your application checks. A simple OpenAI SDK request through 59API can look like this:

client = OpenAI(base_url="https://api.59api.com", api_key=os.environ["59API_KEY"])

response = client.chat.completions.create(model="gpt-4o-mini", messages=[{"role":"user", "content": prompt}], max_tokens=500)

print(response.usage.prompt_tokens, response.usage.completion_tokens)

Use the model identifier supported by your 59API account. The important detail is the measurement: compare completion tokens before and after each prompt change instead of judging only by how short the answer looks.

Replace vague requests with a response contract

“Explain this clearly” gives the model no practical stopping point. Define the audience, output shape, and maximum scope instead:

Bad: Explain this error clearly.

Better: Explain the error for a Python developer in 3 bullet points. Include the cause, one fix, and one prevention tip. Do not include background theory.

This works because the model has fewer reasonable ways to interpret the task. “Three bullet points” controls structure, while the required fields prevent a short but incomplete answer. Avoid adding instructions that do not affect the result; every repeated rule consumes input tokens on every request.

Separate stable instructions from changing data

If your application sends the same policy, formatting rules, or role description with every request, keep that text compact and place the changing document separately. Instead of repeating a long paragraph such as “You are an expert analyst…” on every call, use a concise system instruction:

Return only valid JSON with keys: issue, severity, recommendation. Keep each value under 30 words.

Then send only the ticket, log, or document that needs analysis. Store reusable instructions in your application code and remove duplicated examples unless they solve a known failure. One carefully selected example is often more efficient than five similar examples.

Use the cheapest suitable model

Do not use a high-capability model for every step. A practical routing pattern is to send classification, extraction, rewriting, and simple summaries to a faster Claude Haiku or GPT model, then reserve Sonnet or Opus for ambiguous reasoning and difficult code work. Fable can be selected when it fits the task and is available in your account.

• Use a smaller model for intent detection and field extraction.
• Escalate only when confidence is low or validation fails.
• Use a larger model for a second pass, not as the default first pass.

59API is useful for this workflow because it provides pay-as-you-go access to native official-quality Claude and GPT models through one relay, without requiring a subscription to each provider. Its pricing is designed for low-cost usage, and its referral rebate can further reduce the effective cost for teams with ongoing traffic.

Limit output deliberately

Set a reasonable output ceiling in the API request, but do not treat it as a substitute for prompt design. For a support-labeling task, 500 tokens is excessive. A limit such as 100 or 150 tokens, combined with a fixed JSON schema, makes runaway explanations less likely:

max_tokens=150

Also specify what to omit: “No preamble, no restatement of the input, and no conclusion after the requested fields.” These three exclusions remove common padding without weakening the useful content.

Test quality, not just token count

Create a small evaluation set of real requests and compare the old and new prompts on correctness, missing fields, refusal rate, and token usage. A prompt that saves 40 percent but causes an extra retry is not actually cheaper. Log failed validations and use them to add one targeted rule rather than making the entire prompt longer.

Once the shorter prompt meets your quality threshold, route it through 59API and monitor usage in production. You can sign up for 59API, add pay-as-you-go credit, and connect it to your existing Claude Code, Codex, or OpenAI SDK workflow using https://api.59api.com.

Precios · Docs