Claude vs Steve AI

Two model launches in under four weeks — Sonnet 5 on June 30, Opus 5 on July 24 — with Opus 5 positioned near Fable 5's intelligence at half the price. Around the models the surface area keeps widening: Cowork moving to web and mobile with sessions running remotely, a Microsoft 365 connector that writes instead of only reading, and now memory that spans chat and Cowork alike. The admin track has kept pace with Trusted Devices, model entitlements, self-serve HIPAA, and scanning for third-party skills and plugins.

The pattern through most of this window is reach first, govern second: each capability that puts Claude inside someone's workplace is followed within weeks by a control that constrains it. The memory release breaks that order — cross-surface recall, per-item editing, and a sensitive-topics boundary all land together, with the default set by plan rather than by the user. If that sequencing holds, governance stops trailing capability and starts shipping as part of it.

Expect the same simultaneous treatment applied to Cowork's remote sessions and connector write permissions, where controls are still catching up. Whether skill and plugin scanning leaves Enterprise beta is the clearest near-term signal that the governance track is being generalized rather than sold upmarket.