AI Token Budget Planner

A rough rule: 1 token ≈ 0.75 words or 4 characters. A short question is ~50-100 tokens. A full document summary may use 2,000-8,000 tokens. Count system prompt, conversation history, and output separately. Most providers offer token counting APIs (tiktoken for OpenAI, claude_tokenizer for Anthropic) for accurate estimates.

Output tokens require the model to generate each token sequentially, which is computationally more expensive than reading input. Typically output tokens cost 3-5x more than input tokens. This is why minimizing verbosity in responses has a big impact on cost.

For simple classification or extraction: use Haiku or GPT-4o mini. For complex reasoning, code generation, or long-context tasks: use Sonnet or GPT-4o. For the hardest tasks requiring maximum intelligence: use Opus or o1. Most production apps use a tiered approach — cheap models for simple tasks, expensive models only for complex ones.

Anthropic's Prompt Caching and OpenAI's cached tokens reduce costs by reusing repeated prefix tokens (like system prompts). If your system prompt is 2,000 tokens and you make 1,000 calls/day, caching can save ~90% of input costs on that portion. Factor this into real estimates if you use caching.

A small internal tool: 10-100 calls/day. A production SaaS with 100 users: 500-5,000 calls/day. A high-traffic consumer app: 10,000-100,000+ calls/day. Start with your p50 user and multiply by active users. Add headroom for retries and background tasks.