Lesson 20 / 27

Cost, Latency and Caching

Estimate API cost from tokens and reduce it.

By · 15 years in web development

You pay per token

Hosted APIs charge per million input tokens and per million output tokens, with output usually priced higher. Latency has two parts: time to first token (reading the prompt) and generation speed (tokens per second), and long outputs take longer. To cut cost and delay: shorten prompts, retrieve only relevant chunks, cap output length, use a smaller model for easy tasks, cache repeated prompt prefixes where the provider supports it, and batch non-urgent work. The prices below are example numbers, not any provider's real price list.

A monthly cost estimate, run

I ran this plain-Python (standard library only) example. Ten thousand calls with a 2,000-token prompt and 500-token answer cost $135 at the example prices; caching the prompt at 10% of its price would cut that to $81.

price_in, price_out = 3.0, 15.0     # example prices, USD per million tokens
prompt_tokens, output_tokens, calls = 2000, 500, 10000
cost = calls * (prompt_tokens * price_in + output_tokens * price_out) / 1e6
print("monthly cost: $", cost)
cached = calls * (prompt_tokens * price_in * 0.1 + output_tokens * price_out) / 1e6
print("if the 2000-token prompt were cached at 10% price: $", cached)

Output:

monthly cost: $ 135.0
if the 2000-token prompt were cached at 10% price: $ 81.0

Quick check: Which change usually lowers cost the most for easy classification tasks?

  • Using the largest model with a long prompt
  • Using a smaller model with a short prompt
  • Raising the temperature
  • Adding more examples to every call
Answer

Using a smaller model with a short prompt — Match model size and prompt length to the difficulty of the task.