Understand what you're paying for.
What an LLM feature actually costs, and why: the chain of calls it runs, how the bill reads, what reasoning tokens add, where the tokens go, and which levers cut it. This is the per-token side of pricing, where the cost model has the most moving parts; embeddings, rerank, speech, image, and video models bill in simpler native units, priced on their own category tabs.
How LLM API pricing works
Tokens, the input/output split, cached reads, and why the same answer costs several times more to write than to read.
What drives the cost of common features
RAG, chat, classification, summarisation, a coding agent — each has a different cost shape. Here's why.
Cost-cutting strategies & savings
Tiering, routing, caching, shorter outputs, batch. The levers that move a bill the most, and by how much.
Reasoning & "thinking" tokens
Reasoning models bill the hidden thinking they do before answering, at the output rate, and it is often the biggest line on the invoice. Effort dials the volume of that thinking, not the price per token.
Prompt caching
Reuse a big system prompt or document across calls and pay up to 90% less for the repeated part.
Batch processing
Trade latency for around half off. When a job can wait minutes, the async batch endpoint cuts the bill.
What an AI agent actually costs
An agent makes many model calls per task, each carrying a growing transcript. The cost compounds fast.