Cost Intelligence · AIPersonalization.cloud
What your AI model actually costs
Sticker prices don’t tell you much on their own — output tokens run 4–6x the price of input across every major lab, so a “cheap” model with chatty answers can quietly outspend a pricier one that talks less. Set your own volume below and see the real monthly number for each model.
Rates verified July 28, 2026 · updated when providers change pricingSet your workload
Estimate monthly tokens in millions. Not sure? Pick a preset — these are rough shapes, not universal truths, so adjust once you know your own usage.
million / mo
million / mo
Assume Batch API pricing Async, non-real-time jobs — most providers cut both rates roughly in half
Monthly cost, cheapest first
Reference rates
The raw per-million-token prices behind the calculator above, so you can check the math or model your own scenarios by hand.
| Model | Tier | Input | Output |
|---|
Before you budget off this
What this calculator doesn’t capture
- Prompt caching isn’t modeled. If your workload reuses a stable system prompt or long document, cached input tokens cost 90% less on most providers (and up to 98% less on DeepSeek). Heavy reuse can shrink these numbers a lot.
- Long-context surcharges aren’t applied. Gemini 3.1 Pro roughly doubles its rate above 200K tokens per request, and several GPT-5.6 tiers do the same above 272K. This tool assumes standard-tier pricing throughout.
- Claude Sonnet 5’s rate is introductory. It’s $2/$10 per million through August 31, 2026, then moves to $3/$15 — the calculator uses today’s rate.
- Quality isn’t priced in. The cheapest model per token isn’t automatically the cheapest way to solve a task — a weaker model that needs three retries can cost more than a stronger one that gets it in one pass.
- Prices move fast. Every lab here has repriced at least once in the last few months. Treat this as a same-day planning tool, not a permanent record — check the provider’s own pricing page before signing off on a budget.