A small chunk of text, often part of a word. Providers publish how their tokenisers count text; a rough rule of thumb is a few characters per token in English.

Is output more expensive than input?

With most providers, yes. Output tokens typically cost several times more than input tokens.

Does prompt caching really save money?

For workloads that resend the same long instructions or documents, caching can cut the cost of that repeated portion substantially. Check your provider's caching rules.

blog/llms & ai models

LLM Costs Explained: Tokens, Context and How to Keep AI Bills Under Control

SoftiIntel Team · 30 September 2026 · 6 min read

the short answer

Most LLM APIs charge per token, separately for input (what you send) and output (what the model writes), with larger models costing more per token. Costs grow with long prompts, big documents, lengthy outputs and high volume. The biggest savings come from choosing the right model tier per task, caching repeated context, keeping prompts and outputs lean, and batching work that doesn't need an instant answer.

How pricing works

Input tokens: your instructions, documents and conversation history.
Output tokens: the model's reply. Usually priced higher than input.
Model tier: faster, smaller models cost far less per token than the most capable ones.
Extras: some providers charge for tools like web search, or for reasoning tokens.

Where AI bills balloon

Sending entire documents when a few passages would do.
Resending long conversation history on every turn.
Using the most expensive model for simple tasks.
Verbose outputs nobody reads.
Retries and loops in agents with no limits.

Seven ways to cut cost

Match the model tier to the task and route simple requests to a fast model.
Use prompt caching for repeated instructions and documents where the provider supports it.
Retrieve only relevant passages (RAG) instead of whole documents.
Ask for concise outputs and set sensible length limits.
Batch non-urgent work. Several providers discount batch processing.
Cap agent steps and budgets.
Measure cost per completed task, not per request.

Budget before you build

Estimate volume (requests per day), average input and output size, and the model tier, then multiply by the provider's current rates. Check official pricing pages, because rates change. We include running-cost estimates with every AI agent and integration proposal.