Write Cheaper Prompts: Six Habits That Cut Your LLM Token Bill

Most teams look for savings in the model dropdown. The prompt is usually the cheaper lever. Every call is billed as:

cost_inr = input_tokens / 1e6 x input_price_per_M
         + output_tokens / 1e6 x output_price_per_M

Reasoning models add hidden thinking tokens, which are billed as completion tokens (see reasoning tokens). Look up the live per-million INR prices on the models page and plug your own numbers into the cost calculator. Here are six habits that reduce the variables you control.

1. State the task and the output format once, precisely

Vague prompts produce long, hedged answers. Naming the format ("return JSON with keys label and reason, nothing else") shortens output, which is usually the pricier side of the bill. The clarity and specificity lesson on MasterPrompting.net covers the pattern.

2. Cap the output

Set max_tokens to what the task needs. A classifier does not need 2,000 tokens of headroom. Pair it with a format instruction so the model does not get cut off mid-answer.

3. Trim context you re-send every call

System prompts and retrieved documents are re-billed on every request. Cut boilerplate, dedupe retrieved chunks, and send only the sections the question needs. MasterPrompting's prompt compression lesson goes deeper on techniques and their quality trade-offs.

4. Use provider prompt caching for long, stable prefixes

If most of your prompt is an unchanging prefix, caching can discount the repeated part. Per our pricing docs, Claude cache reads are billed at a fraction of normal input and OpenAI cached input is discounted; check the docs for current multipliers before you count on a number. Our guide to prompt caching explains how to structure prompts so the prefix stays identical.

5. Use few-shot examples sparingly

Examples improve reliability but cost input tokens on every call. Start with one or two strong examples and add more only if an eval shows a gain. See the few-shot prompting lesson for how many are usually enough.

6. Route simple tasks to small models

Classification, extraction and routing rarely need a frontier model. Because AICredits exposes every model behind one OpenAI-compatible key, changing the model is one string. Test the small model on your own examples first.

Measure in rupees, per successful task

A shorter prompt that fails and gets retried can cost more. Log cost per successful outcome, not per call. Use the dashboard's Usage page for per-request INR cost, and set a per-key budget so experiments cannot run away. Need ready-made prompt patterns to start from? MasterPrompting's prompt library has copy-paste templates.