This is the single biggest cost lever. Don’t use GPT-4o for tasks that GPT-4.1 nano can handle. Classification, sentiment analysis, entity extraction, simple formatting — these work great on cheaper models. Reserve the expensive models for nuanced writing and complex reasoning. I route about 70% of my API calls to mini/nano models now.

Every token in your system prompt gets charged on every single API call. I had a 800-token system prompt that I trimmed to 200 tokens without losing any functionality. That saved me roughly 30% on input costs across thousands of calls.

Set max_tokens to limit output length. If you need a one-sentence classification, don’t let the model ramble. A max_tokens of 50 for a yes/no task prevents runaway output costs.

If you’re sending the same or similar prompts repeatedly, cache the responses. OpenAI also offers prompt caching for long system prompts — after the first call, cached prefixes are 50% cheaper on input. For my chatbot, caching saved about 25% of total costs.

OpenAI’s Batch API gives you a 50% discount on all models. The tradeoff is that results come within 24 hours instead of seconds. For anything that isn’t real-time — report generation, bulk content processing, data analysis — this is free money.

In chatbot applications, the entire conversation history gets sent with every message. Long conversations get expensive fast. Implement a sliding window (keep last N messages) or summarize older messages to reduce token count.

Use the OpenAI usage dashboard and set spending limits. I set a hard cap of $50/month and a warning alert at $30. It’s saved me from unexpected spikes more than once, especially during testing when I accidentally left a loop running.