OpenAI, Claude & Gemini API cost calculator

Enable prompt caching as the first optimization, before anything else. For Anthropic, mark your stable system prompt with cache_control headers and cached tokens drop to roughly 10% of standard input price. For OpenAI, caching happens automatically on certain endpoints when your prompt prefix is identical across requests. The win is largest on workloads with long stable context and many requests: chatbots with detailed personas, RAG systems with repeated retrieval context, and any agent loop that re-sends a long instruction set. Most teams forget to enable caching and overpay 5 to 10 times.

Move every non-real-time workload to the Batch API. Both Anthropic and OpenAI offer batch endpoints at exactly 50% of standard pricing in exchange for a 24-hour response SLA. Classification jobs, content moderation, embedding generation, summarization pipelines, and evaluation runs are all good candidates. The integration work is trivial because the prompt and model are unchanged. Teams routinely run their entire content tagging pipeline at standard pricing when batch would cut the bill in half with zero quality difference.

Route requests by difficulty. Simple queries (greetings, formatting, basic lookups) belong on GPT-4o mini, Claude Haiku, or Gemini Flash at $0.15 to $0.80 per million input tokens. Hard reasoning belongs on Claude Opus, GPT-4.5, or Gemini Pro at $1.25 to $75 per million. A small classifier in front of your model router typically cuts costs 60 to 80% on mixed workloads. Sending everything to a frontier model is the most common AI cost mistake we see in production.

Cap max_tokens aggressively. Output tokens cost 3 to 5 times more than input tokens, and the default 4K output buffer is far larger than most responses need. Cap yes-or-no answers at 10 tokens, short summaries at 200, structured JSON at whatever your schema requires plus a small buffer. The model sometimes wants to elaborate even when you did not ask, and you pay for those elaborations. Tightening max_tokens is a 30 to 50% output cost reduction in most cases.

Strip retrieved context to only what is needed per query. RAG systems often retrieve 10 to 20 chunks and stuff all of them into the prompt for safety. The result is paying for thousands of irrelevant tokens per request. Adding a reranker that filters down to the top 3 to 5 most relevant chunks typically cuts input token usage by 50 to 70% with no quality loss, often with quality improvement because the model is not distracted by irrelevant context.

Log every request with token counts, model, and cached versus uncached status. You cannot optimize what you cannot see, and AI costs can change shape overnight when a new feature ships or a prompt gets tweaked. Set daily spend alerts. Build a dashboard that breaks down spend by feature, model, and endpoint. Most production cost overruns we encounter happen because a usage pattern shifted two weeks before anyone looked at the bill.