AI Token Costs Guide

LLM APIs bill in tokens — sub-word pieces — not whole words or characters. English text is often ~0.75 words per token, but code, non-English text, and structured JSON can differ. Providers quote separate per-million rates for input (prompt) and output (completion). Treat published $/1M figures as the unit of comparison across models.

Output tokens almost always cost more than input tokens. Chat and agent apps that stream long answers therefore cost more than classify/extract jobs with short replies. When you model a feature, estimate prompt size and expected completion length separately — using a single “average tokens” number hides the expensive half of the bill.

Long contexts raise cost even when the user message is short: system prompts, tool schemas, retrieved documents, and chat history all count as input. A 100k-token context at a mid-tier rate can dominate spend before the model writes a single sentence. Prefer retrieval that returns only the needed chunks, and cap history windows.

A practical estimate: (input_tokens / 1e6 × input_rate) + (output_tokens / 1e6 × output_rate). Multiply by daily requests for monthly burn. Use the same sample workload (for example 1M input + 500K output) when ranking models so differences come from price, not from inconsistent assumptions.

List price is not the whole story: free tiers, startup grants, volume discounts, and relays change effective cost. Still, you need the token math first — otherwise “cheaper provider” comparisons mix billing units. After you understand tokens, use calculators and ranked lists for decisions.

This page explains the pricing unit. Use /ai-api-cost-calculator or /openai-api-cost-calculator to run numbers, /cheapest-llm-api for a ranked leaderboard, /llm-cost-optimization for reduction tactics, and /api-credits for live model price tables.