AI Models

The models kapi ships support for, with their ceilings and their place in the neokapi lifecycle. This is the committed catalog the binary itself reads — the same file drives the per-provider defaults and the output-token budgeting — so the list here is what your build actually supports, not a hand-maintained copy that drifts.

Each provider leads with the models recommended for the work kapi does — translation, review, terminology — which for most projects is a short list. Advanced & specialised holds models that are fully supported but not the usual choice: capable-but-premium ones (Opus, Gemini Pro) that are overkill for faithful content work, or ones tuned for other tasks (Fable, for creative writing). Legacy holds superseded models — still callable, with a newer successor. Nothing here is hidden or blocked: every model, in any group, can be named with --model, and a model the provider stops serving is removed from the catalog rather than kept as a tombstone.

“In neokapi since” is the date the model id entered the source tree; models present when the catalog was first seeded share that date, and precise dates are recorded going forward. For what these models cost and how they behave under batching, see the batch eval.

Catalog reviewed 2026-07-15. Maintained in providers/ai/models.json; make check-models flags when a provider’s live model list diverges from it.