Last updated: May 2026
Pick your LLM, embeddings, vector DB, hosting, and monitoring — see your total estimated monthly cost live.
Prices last updated: May 2026. Always verify at provider websites. Rates are standard API prices and exclude free-tier credits, enterprise agreements, or promotional discounts. LLM pricing: openai.com/pricing, anthropic.com/pricing, ai.google.dev/pricing.
Cost Summary
This calculator estimates your total monthly cost to run an AI-powered application by modeling each infrastructure component independently. Toggle components on or off to reflect your actual stack — only enabled components count toward your total.
Embeddings have two cost components: a one-time ingestion cost (embedding your document corpus) and an ongoing query cost (embedding each user query). The one-time cost is shown separately and included in Year 1 Total but not Monthly Total.
Vector database storage is auto-estimated from your document count and embedding model dimensions: num_vectors × dimensions × 4 bytes / 1,073,741,824 GB. Query costs are calculated from your daily query volume × 30 days.
Hosting, orchestration, monitoring, and auth are flat monthly fees — choose your plan from the dropdown or enter a custom amount. The Cost per 1,000 requests metric divides total monthly cost by your monthly request volume, helping you understand unit economics for pricing and margin planning.
The monthly cost of running an AI app varies enormously based on the stack you choose. A lean MVP using GPT-4o mini, free-tier services, and Vercel Hobby can run for as little as $2–$15/month. A production-ready stack with Claude 3.5 Sonnet, Pinecone, and Railway Pro typically costs $80–$300/month depending on usage. Enterprise-scale deployments with high request volumes can exceed $500–$5,000/month. The biggest cost driver is almost always LLM inference — especially input/output token volume.
The cheapest viable AI stack for an MVP typically combines GPT-4o mini or Gemini 1.5 Flash for inference (both under $1/million tokens), OpenAI text-embedding-3-small for embeddings ($0.02/1M tokens), Pinecone Starter (free tier) or pgvector on an existing database for vector storage, Vercel Hobby for hosting ($0), and Langfuse free tier for monitoring. With 100 requests/day at 1,000 input tokens and 500 output tokens, GPT-4o mini costs under $5/month for inference alone. Total stack cost can realistically be $2–$15/month.
You need a vector database if your AI app uses Retrieval-Augmented Generation (RAG) — i.e., it searches through your own documents, knowledge base, or data to answer questions. If you're building a pure chatbot, code assistant, or summarization tool that doesn't retrieve from a custom corpus, you can skip the vector DB entirely. For RAG use cases, self-hosted options like Chroma or pgvector cost nothing extra if you already have a server, while managed services like Pinecone start free and scale by storage and query volume.
LLM inference is almost always the dominant cost in an AI stack, often accounting for 70–95% of total monthly spend. This is because inference costs scale directly with usage — every request burns input and output tokens. At scale, a production app sending 10,000 requests/day with 2,000 tokens each to GPT-4o costs $1,500/month just for inference. Embedding costs are usually 10–50x cheaper than inference. Vector database, hosting, and monitoring costs are typically flat or lightly usage-based and rarely exceed $50–$200/month for most apps.
Key strategies to reduce AI API costs: (1) Use a smaller model for simpler tasks — GPT-4o mini is 17x cheaper than GPT-4o. (2) Enable prompt caching for repeated system prompts — saves up to 90% on cached tokens with Claude, 50% with GPT-4o. (3) Reduce output token length by prompting for concise answers. (4) Cache common responses at the application layer to avoid repeat API calls. (5) Use batch APIs for non-real-time jobs (50% discount from OpenAI/Anthropic). (6) Switch to self-hosted open-source models like Llama 3 via Groq for high-volume workloads. (7) Pre-filter retrieval so the LLM sees fewer, more relevant tokens.