Skip to main content
Estimate your agent’s inference cost by choosing a workload profile or entering your own.

The interactive calculator needs JavaScript, which your browser doesn’t support. Here is a representative worked example instead: a deep-research agent that runs 50 turns and totals 4.0M cached input reads, 1.5M fresh input tokens, and 250K output tokens, priced on Sail’s zai-org/GLM-5.3 and on generic traditional inference provider tiers at generic list prices (Sonnet-class: $3 input / $15 output / $0.30 cache reads per 1M tokens; Opus-class: $5 / $25 / $0.50).

​
How the math works

  • Fresh input: tokens the model reads for the first time each turn (new tool results, search snippets, file contents).
  • Cached input: tokens reread from prompt cache (the growing conversation history). Cache reads are billed at the cached rate.
  • Supercache reads: tokens read from Supercache. They use the Supercache read rate.
  • Supercache writes: tokens written to Supercache. They use the Supercache write rate.
  • Output: tokens the model writes (reasoning and answers).
Sail rates come from the Pricing page for the selected model and completion window.

​
Assumptions and caveats

  • Each turn must fit the model’s supported token budget. The calculator counts all input categories plus output per turn. If that total exceeds the selected model’s budget, it shows a warning instead of cost estimates and keeps your inputs unchanged.
  • Traditional inference provider tiers use generic list prices. “Sonnet-class” is $3 input / $0.30 cache reads / $15 output per 1M tokens (batch tier = 50% off inside a 24-hour window); “Opus-class” is $5 / $0.50 / $25.
  • Sail’s ASAP prices undercut traditional inference providers for many models. The calculator uses Sail’s published rates and generic traditional inference provider list prices. A same-model row appears only when the pricing data includes a maintained traditional inference provider price. The calculator does not substitute ASAP for that price.
  • The model has to do the job. For simplicity, the math assumes you’re using a single frontier-class open model for your task. Often, we see the hybrid approach using both frontier closed models and open models, or a mix of open models.