7 scenarios
What it actually costs
“It depends” is the honest answer and a useless one. It depends on the token shape of the work — so each of these states a shape, then computes the bill from it across models and volumes.
Nothing here is a hardcoded price. Every figure is calculated at build time from the 42 verified rates in the price table, last reviewed August 5, 2026.
Chat
Retrieval
Automation
- An agent that uses tools38,000 in · 3,000 out · per taskhow much does an ai agent cost
- Classifying support tickets900 in · 60 out · per ticketllm cost for text classification at scale
- Processing documents at scale22,000 in · 700 out · per documentcost to process documents with llm
- A code assistant in your editor11,000 in · 1,400 out · per interactionai coding assistant cost per developer
Content
Why most cost estimates are wrong
Almost every budget for an LLM feature is built by taking the price per million tokens, multiplying by a guessed prompt length, and multiplying by request volume. That calculation is wrong in the same three ways every time.
It measures the user’s message rather than the assembled prompt, which also carries the system instructions, the retrieved context, the tool schemas and the conversation history — routinely five to twenty times larger. It ignores that conversations and agent loops resend their history, so cost grows faster than request count. And it prices at the uncached rate while assuming caching that was never switched on.
What these pages do instead
Each one states the token shape of a single unit of work — a conversation, a document, an agent task — and says where those figures come from. Then it prices that shape across three models and three volumes, showing what caching is worth on that specific workload rather than in general.
The savings sections are ordered by return per hour of effort, and each one names the mechanism rather than the aspiration. “Enable prompt caching, roughly 40%” is checkable. “Optimise your prompts” is not.
The shapes are representative, not measured
They come from what these workloads typically look like, and each page says so in the shape block. Yours will differ, sometimes by a lot — which is exactly why every page ends by pointing at the cost calculator and the token counter, where you replace the representative numbers with your own.
If you would rather work from the rates directly, the price table carries all 42 models with the provider URL each figure was read from, and the cost control checklist is the twenty checks that cut a bill without changing what users get.