August 11, 2026
GPT-5.6, Kimi K3, and GLM 5.2 join your roster.
Five new models for your teammates. GPT-5.6 Terra arrives with its heavier and lighter siblings Sol and Luna, alongside Kimi K3 from Moonshot AI and GLM 5.2 from Z.AI. Every one cleared the same bar before it got in: long real tasks, zero silent failures. Pick a model per teammate in the picker or from the CLI, and switch any time.
New
- Five models join the roster: GPT-5.6 Terra, Sol, and Luna from OpenAI, Kimi K3 from Moonshot AI, and GLM 5.2 from Z.AI.
- Sol and Luna are Terra's heavier and lighter siblings: the same 1M context, tuned up for the hardest work or down for the quick stuff.
- Every model cleared the same bar before joining: long real tasks, zero silent failures.
- Pick the model per teammate in the model picker, or dock hire --model from the CLI. Switch whenever you like.
- New billing plans: Starter at $20 a month, Standard at $100, Scale at $200. When your usage runs out the choice appears in the conversation instead of a settings page, and your teammate picks the work back up once the payment clears.
- Vault keys now open per agent: a new key starts locked, so nobody can use it until you add them to that key's list, and the in-chat add-key card grants the asker as you add. Keys you already had show which teammates were auto-added from earlier use, so you can trim the list.
- Queued messages are a drawer: see what waits behind a busy agent, send one ahead, dismiss one, or clear the backlog. Stops land in seconds.
Polish
- The model picker got cleaner: current models up front, the rest behind a More models menu.
- Replies from the new models stream smoothly, with real activity shown while they work.
- Agent panels open instantly and appear whole. No skeletons, no half-loads.
- Table tabs switch instantly, and the row cache stays honest against your edits.
- Files: folders you just visited stop re-skeletoning, and image previews open cleanly on top.
Fix
- Messages in the drawer are attributed by author, so a cross-org message never reads as You.
- The first-run greeting renders at the same size as every other agent message.
- Delete confirmations use a real in-app modal instead of a browser prompt.
- Workspace cards no longer freeze at a stale width, and docs no longer double-load on switch.
API
- send_message refuses agent recipients outright instead of stranding a message silently.
- request_capability_key: a tool any agent can call to summon the Vault add-key card into the conversation.
- Agents can discover the files surface over MCP, with honest capability info.