KaziKazi

How do AI agent credits work, and what does one message cost?

A credit is a fixed block of model usage, so what a message costs depends on how much reading and writing the model does for it. In Kazi one credit is 100,000 tokens priced at GPT-5.6 Luna rates, with output counted six times and cached input a tenth. A typical message uses well under one credit.

Ken Wiltshire, founder of Kazi · Checked against the product on 2026-11-14

Why tokens, not messages

"Hi, what's on my calendar?" and "read these 40 contracts and flag the renewal terms" are both one message. They are not the same amount of work. Pricing per message either overcharges the first or loses money on the second.

Tokens track the actual work: what the model reads, including the files and tool results it pulls in, and what it writes back.

How Kazi counts

  • 1 credit = 100,000 tokens at GPT-5.6 Luna rates.
  • Output tokens count 6×, because writing costs the model more than reading.
  • Cached input counts 0.1×. In a long thread most of the context is cached from the turn before, so continuing a conversation is cheap.
  • Pricier models use credits faster, cheaper ones slower. The rate is the same.

The same balance works on cuttingfloor, where one credit is one render minute. Turns on your own model API key don't touch it.

What makes a task expensive

Big inputs. An agent that reads a 300-page PDF into context pays for every page. Kazi can run code in a sandbox for that kind of job, so it searches and slices the file instead of reading all of it into context.