LLM cost calculators by workload

Generic price tables hide the thing that actually decides your bill: the ratio of input to output tokens, and how much of your input repeats. These eight calculators fix that ratio to a real workload shape and rank every model accordingly.

• Customer support chatbot
  A support assistant that answers from a system prompt plus a few retrieved help-centre snippets, then replies in a short paragraph. The system prompt is identical on every call, so prompt caching does most of the work here.
• RAG document Q&A
  Retrieval-augmented Q&A over a private corpus: several thousand tokens of retrieved chunks per question. Input dominates cost, and retrieved chunks change per query so cache hit rates stay low.
• Coding agent
  An agent that reads a repository, plans, and writes patches across multiple turns. Very large inputs, heavy re-reading of the same files, and long outputs — the single most cache-sensitive workload on this site.
• Bulk summarization
  Summarizing a large backlog of documents, tickets, or transcripts. Almost pure input cost, essentially no cache reuse, and output length is tightly bounded.
• Structured data extraction
  Pulling structured fields out of unstructured text at high volume. Individually cheap, but the call count is what makes model choice worth several thousand dollars a month.
• Long-form content generation
  Drafting long articles, reports, or descriptions. The only common workload where output tokens dominate — which flips the usual ranking, because output is priced 4-6x higher than input almost everywhere.
• Email triage agent
  Classifying, routing, and drafting short replies to inbound mail. Short outputs, repetitive instructions, high volume — a workload where budget models are usually the correct answer.
• Translation pipeline
  Machine translation at scale, where output length roughly matches input length. Balanced in/out ratio makes the blended rate the number that matters.