LLM Tool Cost Calculator

Because vendors meter different things. Datadog bills per LLM span, AgentOps bills per event, W&B Weave bills per GB ingested, Pydantic Logfire bills records, Confident AI bills GB-months, Langtail bills the number of prompts you have, OpenRouter takes a percentage of provider spend, and LangSmith bills seats. A published price of $50 a month means something completely different in each case. This calculator normalises them to one workload so the numbers are actually comparable.

Because it is the difference between a manageable bill and an unmanageable one, and it is the single most common estimating mistake in this category. A simple completion produces roughly 3 to 5 spans. An agentic workflow with tool calls, retrieval and reasoning steps commonly produces 20 to 50. Any tool that meters spans therefore charges an agent workload up to ten times what it charges a completion workload at identical request counts. Datadog's 40,000-span free tier is a few thousand simple requests, or a day or two of agent traffic.

Because several vendors bill volume rather than events. W&B Weave charges per GB ingested and Confident AI charges GB-months, which means two applications making identical numbers of calls can have very different bills. A RAG system logging a long system prompt, ten retrieved chunks and a lengthy completion produces a far larger payload than a classification endpoint returning one word. If you bill by volume, verbosity is the cost driver.

Treat them as a modelling tool rather than a quote. They are computed from published rates we verified against vendor pricing pages, but real bills depend on negotiated terms, annual commitments, regional pricing and tier boundaries we cannot see. Several vendors publish no pricing at all and are marked as such rather than guessed at. Use this to find the shape of the answer and to spot the tools that are an order of magnitude wrong for your workload, then confirm with the vendor.

Either they are open source and self-hostable at no licence cost, or your workload fits inside their free allowance. Free on licence is not free to operate - self-hosting Langfuse, SigNoz or LiteLLM means running the infrastructure, which is real money and real engineering time even though no vendor invoices you. The calculator shows licence and subscription cost, not total cost of ownership.

Judge model inference, which is frequently the largest hidden cost. Evaluation frameworks like Ragas and guardrails like NeMo make additional model calls for every check they run, and those land on your model provider invoice rather than on the tool vendor's bill. A thousand-case eval suite across four metrics is several thousand judge calls per run. Self-hosting infrastructure is also excluded, as is engineering time.