How Chat Turns Work
What happens on each turn of a chat, how tokens are counted and how the price is worked out.
A turn, step by step
A model remembers nothing between calls. Each time you send a message, your app sends the whole conversation again, and the model reads it from the start. One request and its reply is a turn. Press play to follow five turns of a chat.
Press play, or step through one stage at a time.
Request
Nothing sent yet.
Usage history
No turns billed yet.
Sila groups the turns into one conversation in Usage History, but prices each turn on its own.
Why each turn costs more
The history is sent again on every turn, so the input grows even when your messages stay short. By turn 5 you have typed 10 tokens, but the model reads 1,542.
- Start a new chat when the topic changes.
- Keep system prompts and tool lists short; they ride along every turn.
- Ask for short answers when you can; each reply joins the history.
Counting tokens
A token is a piece of text, often a short word or part of a longer one. Sila does not count tokens itself. It takes the counts the vendor reports for each call, so they match the vendor's bill.
Input tokens
- The system prompt, yours and the model's in the catalog
- Every earlier message and reply
- The new message, with any images or files
- Tool definitions and tool results
Output tokens
- The reply text
- Thinking, for models that reason
- Tool calls the model makes
Try it
about 21 tokens · 60 characters
An estimate for learning. Thai and other languages without spaces usually take more tokens than English for the same meaning.
Working out the price
Every model has an input price and an output price, per 1 million tokens. Sila prices each turn on its own:
+ output tokens × output price ÷ 1,000,000
- A call that fails costs nothing.
- Charges are kept to 8 decimal places.
- Cached input has no price of its own yet; it is billed as input.
- When a hard budget runs out, Sila refuses the next call before it reaches the vendor.