Help for owners
My AI usage bill keeps growing and I do not know why
Short answer
AI bills grow with the amount of text sent and received, how often calls happen and which model is used. Today, set a spending limit in the provider's dashboard, check who holds the keys, and look at usage by day. Then reduce waste by trimming prompts, reusing repeated text and using smaller models where they suffice.
Does this sound like your situation?
- The monthly bill jumped without a matching rise in users.
- I cannot tell which feature causes the cost.
- A key may be shared or exposed.
- The same long instructions are sent with every request.
- One user or script seems to use a lot.
Why does this happen?
Every request sends the whole prompt
Long instructions and documents are sent again with each call, so they are paid for each time.
Repeated text can be cached
Anthropic documents prompt caching, which reuses an identical prompt prefix; cache reads are billed at a fraction of the normal input price, though it needs identical content and a minimum length.
Confirmed by platform.claude.com
The most capable model is used for every job
Simple tasks may not need the biggest model. Which one is enough is found by testing on your own questions.
Nothing limits abuse or loops
Without per-user limits, one script or a bug that retries can use a lot, and providers also apply rate limits.
Technical detail: The OpenAI API answers 429: out of credit, or just too fast?
What can I do about it today?
- Set a spending limitIn your provider's billing settings, set a monthly cap or alert, if the provider offers it.
- Look at usage by dayMost providers show usage per day or per key. Note the day the cost jumped and what changed then.
- Check who has the API keyIf it was shared in chat or put in an app, create a new one and remove the old after your developer swaps it.
- Do not delete the key blindlyRemoving a key in use will stop the feature. Plan the swap with your developer.
- Keep the account in your company's nameSo you can see invoices and control access.
When should I bring someone in?
- The bill is much higher than expected and you cannot find why.
- You suspect a key has leaked.
- Cutting cost seems to reduce answer quality.
How do I handle it?
- Find where tokens goI log requests by feature and user to see what each costs.
- Trim and reusePrompts are shortened, and repeated parts are cached where the provider supports it.
- Match model to taskTests on your questions show where a smaller model gives the same quality.
- Add limitsPer-user and per-day limits and retry caps stop loops and abuse.
- Track itA simple report shows cost per feature so changes are visible.
The price depends on what I find, so I quote in writing after a short discovery call. Tell me what is happening and I reply within 24 hours.
What do owners ask about this?
Will cheaper settings make answers worse?
Sometimes. That is why I test each change on your own questions before switching.
Is caching available with every provider?
No. Features differ by provider, so I check the provider's documentation.
Can I put a hard cap on spend?
It depends on the provider. Many offer limits or alerts in billing settings.
Sources
How was this page made?
I drafted this page with AI assistance from the official documents listed under Sources. Found a mistake? Tell me.