Hard Token Cap

How it works

Large language models process text in units called tokens — roughly 0.75 words per token in English. A typical support conversation consumes 500–2,000 tokens depending on length, complexity, and how much of your knowledge base the AI retrieves.

Vendors who charge per-token impose hard caps to control infrastructure costs. When a conversation hits the cap, the AI stops generating output. In a support context, this means a customer mid-conversation suddenly receives no response, or a generic 'please contact us' fallback — at exactly the moment they need help most.

Hard caps are often buried in pricing pages as 'included tokens' or 'conversation credits.' A 1,000-token monthly cap sounds generous until you realize a single BFCM traffic spike can exhaust it in hours. Some vendors charge overage fees per additional 1,000 tokens, adding unpredictable costs during peak periods.

When you'll encounter it

You'll encounter hard token caps when comparing AI support pricing pages, especially from newer AI-native vendors that expose underlying LLM costs directly. Watch for terms like 'conversation credits,' 'AI credits,' 'included tokens,' or 'automation sessions' with monthly limits — these are token caps with different names.