LLM spending should be treated as an operating risk, not an afterthought discovered when the invoice arrives. The post says token-based pricing behaves differently from ordinary cloud costs because it depends on the changing shape of conversations.
A feature can look cheap during a demo, then become expensive when users paste long documents, sessions grow, or an agentic loop calls a model several times for one user action. Retries after malformed responses can add more hidden cost without any obvious product change.
The problem becomes harder when teams use multiple providers, self-hosted models, and separate dashboards with different definitions of a token. Cloud billing can also lag, so alerts may arrive only after the money has already been spent.
Mozilla’s practical recommendation is to manage LLM costs with the same prevention, detection, and mitigation mindset used for other risks. Teams need visibility before launch, live monitoring during use, and fallback options that do not require turning off a whole feature.