Kobi Kredi

Language model costs are token-based and stay opaque to users: nobody knows in advance how many tokens a question will consume. Kobi Kredi reduces that variability to one unit and makes consumption visible.

There are two core metering points. Document upload has a fixed cost; an AI message starts from a minimum threshold and is then computed from actual token consumption. Output tokens are weighted more heavily than input, because provider pricing works the same way.

The result is rounded to one decimal place. That keeps a user’s balance readable and removes the burden of tracking fractions.

Each plan defines a monthly Kobi quota. The quota is independent of user count; a team with few heavy users and a team with many light users consume the same quota at different rates.

This produces a structurally different cost curve from per-seat pricing. Where headcount is large but usage is sparse, the Kobi Kredi model is advantageous; where headcount is small but usage is intense, the balance shifts.

That is why plan selection should rest on pilot measurement rather than theoretical calculation. Two weeks of real usage tells you more than any estimate table.

Four variables drive consumption: the number of passages retrieved, their length, the length of the generated answer, and whether conversation history is added to context. The first two relate to scope and chunking, the last two to usage habits.

The most effective saving is usually narrowing scope. An over-broad assistant retrieves more passages and lowers precision; narrowing improves cost and quality at once.

The second lever is conversation length. Starting a new chat for a new topic is noticeably cheaper than continuing a long one, because history is not resent with every message.