The formula
Monthly cost
monthly cost = requests × (input_tokens × input_price + output_tokens × output_price) ÷ 1,000,000
Example: 100K requests/month at 2,000 input + 500 output tokens on a $2.50/$15 model
100,000 × (2,000 × $2.50 + 500 × $15) ÷ 1M
= 100,000 × ($0.005 + $0.0075)
$1,250/month
The four corrections that make estimates real
- Tokenize your actual payloads. English ≈ 4 chars/token is a rough floor; code, JSON and non-English run heavier. And tokenizers differ by family — Anthropic's newest models produce ~30% more tokens for identical text than older Claude models.
- Count the invisible tokens. System prompts, tool definitions (hundreds of tokens each), retrieved context and conversation history all bill as input on every request.
- Model caching. If most requests share a large prefix (agents, chatbots with long system prompts), cache reads at ~0.1× input price change the math by an order of magnitude.
- Check the deprecation calendar. A model retiring mid-quarter (see the tracker — six OpenAI shutdown waves are scheduled for H2 2026) forces a migration whose replacement may price differently — or, like Sonnet 5 on Sept 1, the same model simply gets +50% dearer.
Run your numbers in the calculator — it uses only provider-verified prices.