Cohere vs GravityKit

Cohere's changelog is almost entirely model announcements: Rerank 4.0 and Command A Reasoning in May, Command A+ later that month, North-Mini-Code in June, Transcribe Arabic in July. Two threads run through them — reasoning and agentic capability on the Command line, and a widening set of specialised models around it for ranking, code and speech. Running alongside is a steady retirement programme that hard-fails requests to older model IDs.

The portfolio is splitting into a reasoning-heavy flagship line and purpose-built models beneath it, with the newer releases sized in explicit tiers rather than offered as one endpoint. Rerank 4.0's pro and fast variants show the pattern clearly: quality and latency become a customer choice instead of a vendor compromise. The deprecation cadence suggests Cohere is willing to keep the supported surface small, which keeps the catalog current but puts migration work on customers on a regular schedule.

Expect the tiered variant pattern to spread to more of the model line, and expect another retirement notice with a hard cutoff date as the newer generations settle. Several recent entries are captured with titles only, so the specifics of the Command A+ and North-Mini-Code releases are not visible here.