Bots aren't free to run. Every time a bot thinks, it uses a model, and models are billed by use. Managing bots includes managing that cost.
How pricing works
Models are priced per token, roughly a word or part of a word. You pay for:
- Input tokens: everything on the desk the model reads.
- Output tokens: what it writes back.
- Cached input: repeated material it has already seen, often at a lower rate.
Here's a real example from xAI's published pricing (docs.x.ai), per million tokens:
| Model | Context window | Input | Cached input | Output |
|---|---|---|---|---|
| grok-4.7 | 500k | $2.00 | $0.50 | $6.00 |
| grok-4.3 | 1M | $1.25 | $0.20 | $2.50 |
Those are the short-context rates. On these models, once a prompt reaches 200,000 tokens, the whole request bills at the long-context rates, which are double ($4.00 input and $12.00 output for grok-4.7). Prices change, so check the page for current numbers.
Why long sessions cost more
Each time the bot responds, it re-reads what's on the desk. A long conversation means a lot of input tokens every turn. That's one more reason to keep the desk clean and put standing material in files. See Context windows.
xAI's docs also recommend a cache key so a conversation's repeated material can be read at the cheaper cached rate (docs.x.ai). That's a setup detail for builders, but it shows how much the repeated part matters.
Picking the model
Not every job needs the biggest model. Think of it like staffing:
| Job | What it needs |
|---|---|
| Sorting, tagging, simple summaries | A fast, cheaper model |
| Drafting, research, judgment calls | A strong general model |
| Hard reasoning, long documents, code | The strongest model, with more reasoning |
Some models also let you set how hard they think. xAI lists low, medium, high and xhigh reasoning settings for Grok 4.7 (docs.x.ai). More thinking usually means better answers on hard problems, and more tokens.
Cost is a resource, like hours
In ISO terms, this is resources: the organization provides what the work needs. Model time is a resource, like labor hours or machine time. You plan it, you watch it, and you don't waste it.
Where cost hides
- Long chats that never get a fresh start.
- Giant pasted documents when one page would do.
- Retries from unclear instructions. A vague job gets done three times.
- Chains of bots that pass whole conversations instead of short files.
Most cost savings come from clearer work, not cheaper models.
What this means for you
- Know which model each bot uses and roughly what it costs.
- Match the model to the job.
- Keep conversations short and standing material in files.
- Check the published pricing page, not memory. It changes.
The cheapest bot is the one that does the job right the first time.