4. Model examples of operations

The calculations below assume CDX-GPT-MAIN in the Standard class with standard context and CDX-EXTRACT in Standard mode.

Credit consumption does not depend solely on the length of the response. While working, the AI has the entire content of CODEXIS at its disposal and, based on your request and the assistant's instructions, browses and reads as many source materials as it needs to be sufficiently certain that it has found exactly what was requested.

This leads to a simple rule: the more precise the request, the less searching, fewer operations and usually a cheaper output. Conversely, a broad or vague request tends to be more expensive, because the model has to go through more content. Formulating a request takes some experience; the Prompting guide for CODEXIS AI is also helpful.

4.1 Case law research (58,635 credits / operation)

Typical course: the EXTRACT model goes through thousands of pages of case law (a large cached context of court decisions), and the MAIN model synthesizes the output (~50,000 output tokens, i.e. roughly 35+ pages of analysis).

Breakdown of tokens and credits per operation:

MAIN token ratio (in:cached:out) = 4 : 50 : 1 EXTRACT token ratio (in:cached:out) = 4 : 4 : 1

4.2 Legal research without case law (17,423 credits / operation)

Less extraction work, a shorter output (~15,000 output tokens, i.e. roughly 10 pages).

Breakdown of tokens and credits per operation: