Reduce your LLM API costs by up to 60%. Strategies: prompt caching, model routing, token reduction, batch processing, and provider negotiation. With real numbers.
Most teams spend 30-60% more on LLM APIs than necessary. Here are seven proven strategies to cut costs. First, enable prompt caching. OpenAI, Anthropic, and Google all offer 90% discounts on cached input tokens. If your app sends the same system prompt with every request, caching alone can cut input costs by 80-90%. For a GPT-5.5 app spending $1,000/month on input tokens, that is $800-900 in monthly savings. Second, use model routing. Send simple queries (classification, short answers) to budget models like GPT-4o Mini ($0.15/1M input) and reserve premium models like GPT-5.5 ($5/1M) for complex tasks. A typical workload can route 70% of queries to cheaper models. Third, reduce output tokens. Add instructions like "respond in under 50 words" or "output JSON only" to cut output costs, which are 3-6x more expensive than input. Fourth, use batch APIs. OpenAI and Anthropic offer 50% discounts for non-urgent batch requests with 24-hour turnaround. Fifth, compress your prompts. Remove redundant instructions, use abbreviations, and trim few-shot examples. Sixth, set max_tokens limits to prevent runaway outputs. Seventh, monitor usage with dashboards to catch cost anomalies early. The fastest win is usually prompt caching — it requires no code changes beyond adding a cache header to your API calls. Model routing comes second: implement a lightweight classifier that scores query complexity and routes to the appropriate tier. Start with these two strategies and you can typically cut your LLM bill by 40-60% within a week. Track your savings by logging per-request token counts and costs before and after each optimization.
Up to 90% on input token costs. If you send the same system prompt repeatedly, caching eliminates most input charges. For GPT-5.5 at $5/1M input, cached input drops to $0.50/1M — saving $4.50 per million tokens.
Model routing sends each query to the cheapest model that can handle it. Simple tasks go to GPT-4o Mini ($0.15/1M), complex tasks go to GPT-5.5 ($5/1M). Most workloads can route 60-80% of queries to budget models.
Yes, significantly. Output tokens cost 3-6x more than input. Cutting average output from 500 to 200 tokens can reduce total cost by 40-60% for most workloads. Add explicit length instructions to your prompts.
OpenAI and Anthropic offer batch endpoints at 50% discount. Requests are processed within 24 hours. Ideal for non-urgent tasks: data enrichment, bulk classification, report generation.
Track daily token usage per model using provider dashboards. Set billing alerts at 80% and 100% of your monthly budget. Log request-level costs to identify expensive query patterns and optimize them.
Output tokens. They cost 3-6x more than input tokens but many teams focus only on input pricing. A model with cheap input but expensive output can cost more in practice than a balanced model. Always calculate total cost including both input and output for your actual usage pattern.