🇬🇧EN
Compare real-world monthly token spend across top providers with prompt caching and batch pricing applied.
Token pricing varies dramatically between frontier reasoning models and high-throughput MoE architectures. This calculator computes real-world spend across DeepSeek, Anthropic, Google, GLM, and OpenAI, accounting for prompt caching and batch discounts.
Open the calculatorApplies to stable input such as system prompts and retrieved context.
50% off for work that tolerates a 24-hour SLA.
Estimated monthly cost
0–0
Median estimate: $25
Calculators give a range. A scoping call gives a fixed scope, timeline and budget.
Talk to our teamInput and output tokens are priced dynamically from our central model registry based on published per-million token rates, then adjusted for prompt caching and batch discounts where supported.
Frontier models cost several times more per token than their mid-tier siblings. Send routine classification and extraction to a small model and reserve frontier calls for reasoning that fails otherwise; on mixed workloads this typically cuts spend by well over half.
Cached input is dramatically cheaper, but only for the stable portion, usually a system prompt plus retrieved context. It pays off when that prefix is consistent across requests, which describes most production RAG systems.
Output usually costs more per token than input, so verbose models get expensive fast. Constrain response length before reaching for a cheaper model.