ad
million
million
| Your cost/mo ▲ | Model | Provider | Input /1M | Output /1M | Blended /1M |
|---|
How we rank. We cost your actual workload: input tokens × the input rate plus output tokens × the output rate. That matters because output usually costs 3-6× more than input, so a model that looks cheap on input can be dear if you generate a lot. "Blended /1M" assumes a typical 3:1 input:output mix for a quick single-number comparison. All rates are standard USD per million tokens, converted live to your currency, batch APIs are ~50% cheaper and prompt caching can cut cached input up to 90%. Prices move fast; confirm on the provider's site.
ad
Common questions
What's the cheapest LLM API?
Proprietary: Gemini Flash-Lite and GPT-4.1 Nano near $0.10/1M input. Cheapest overall is DeepSeek (~$0.14 in / $0.28 out). The best pick depends on your input:output ratio, use the calculator.
Why do input and output cost different amounts?
Generating output is far more expensive than reading input, so output is usually 3-6× the input price. That's why a single blended price is misleading.
Is Claude or GPT cheaper?
Close at the frontier. Claude Opus ~$5/$25, GPT-5 ~$2.50/$15 standard (much more for Pro). Claude's prompt caching can cut input up to 90%, often making it cheaper in practice.
How do I cut LLM costs?
Use a smaller model where you can (budget models are 20-50× cheaper), cache repeated prompts, use the batch API for non-urgent jobs, and trim prompt and output length.