Why AI API bills spike, and how to catch it within a day
AI API spend is the fastest-growing and least predictable line on most small teams' bills. One prompt change can multiply it overnight. Here's what usually causes spikes and how to see them early.
Common causes
- Loops and retries: an agent or job that calls the model again on every error.
- Growing context: conversation history or retrieved documents added to every request, so each call gets more expensive.
- A model switch: moving a feature to a larger model multiplies cost per token.
- Traffic: a launch, a bot, or a scraper hitting an endpoint that calls an AI API.
- A leaked key: keys found in public code get used quickly, and the bill is yours (example).
- Experiments: a teammate's batch job or a weekend prototype on the production key.
Detecting a spike
Look at each model separately, every day. Total spend can hide a spike: if one model triples while another drops, the total barely moves.
A rule that works without tuning: compare each day with the median of the previous 14 days. Flag days above 2.5× the median, and treat 5× as critical. Ignore tiny amounts (under $1) so new or rarely used models don't cry wolf. Burnstack applies exactly this rule, per vendor and per model, from your billed cost.
Limiting the damage
- Set spend limits per project (OpenAI) or workspace (Anthropic).
- Use one project or workspace per app and environment, so you can see, and cut off, exactly what spiked.
- Set
max_tokensand cap retries in code. - Use prompt caching for long, repeated context.
- Keep secret keys on the server; never in a mobile app or front end.
When it happens
- Find the model and project or workspace that spiked.
- If it isn't explained by real traffic, rotate the key first and investigate second.
- Check logs for loops, retries or unexpected callers.
- If it was abuse, contact the vendor's support; some have waived charges from leaked keys.