---
title: "A brain per teammate: GPT-5.6, Kimi K3, and GLM 5.2 join the roster"
excerpt: "Five new models for your teammates — GPT-5.6 Terra with its heavier and lighter siblings Sol and Luna, Kimi K3 from Moonshot AI, and GLM 5.2 from Z.AI. Pick per teammate, switch any time."
author: naomi
category: Launch
date: "2026-08-14"
updated: "2026-08-15"
image: "/blog-covers/a-brain-per-teammate-cover.png"
---

<!-- Product-update post. All claims verified against origin/main: changelog entry "more-models" (Aug 11), ModelPicker.tsx, ModelPickerShot.tsx. Exact facts: 5 models — GPT-5.6 Terra/Sol/Luna (OpenAI, same 1M context, Sol heavier / Luna lighter), Kimi K3 (Moonshot AI), GLM 5.2 (Z.AI). Bar: long real tasks, zero silent failures. Per-teammate picker + dock hire --model. No pricing/availability specifics beyond changelog. -->

Your teammates just got a choice of brains. Five new models join the roster this week: **GPT-5.6 Terra** from OpenAI, alongside its heavier and lighter siblings **Sol** and **Luna** — same 1M context, tuned up for the hardest work or down for the quick stuff — plus **Kimi K3** from Moonshot AI and **GLM 5.2** from Z.AI.

## One team, different minds

Here's the thing this release is actually about: [a teammate is a role, not a model](/blog/who-is-an-ai-teammate). The name, the memory, the address, the routines, the working relationships — all of that belongs to the teammate. The model is the brain it runs on, and there was never a good reason for every role on your team to run on the same one.

So the model is now a per-teammate decision, made where you'd expect: a picker on the teammate, or [`dock hire --model` from the terminal](/blog/your-terminal-into-dock). Your researcher can run heavy while your triage teammate runs light. And it's not a commitment — switch any time, and the teammate stays itself: same memory, same role, same standing work, new brain.

<video autoplay muted loop playsinline aria-label="The model picker in motion: the menu opens over the composer and a different model is selected — the composer pill updates to match.">
  <source src="/blog-assets/a-brain-per-teammate/model-picker-clip-sonnet.webm" type="video/webm" />
  <source src="/blog-assets/a-brain-per-teammate/model-picker-clip-sonnet.mp4" type="video/mp4" />
</video>

*Switching in real time: pick a different brain and the pill updates. Same teammate, same memory, new model.*

## The bar for joining the roster

Not every model gets in. Every one of the five cleared the same bar before joining: **long real tasks, zero silent failures**. Teammates work unattended — a model that quietly drops the ball halfway through an hours-long job isn't a cheaper option, it's a hiring mistake. The roster is curated for the way teammates actually work: long-running, multi-step, nobody watching.

## How to think about the choice

The honest guide is the same one you'd use for people: **match the mind to the work**.

- **The hardest lanes** — deep research, gnarly refactors, anything where being wrong is expensive — deserve the heavier end: GPT-5.6 Sol, or the Claude models your roster already runs on.
- **The high-volume lanes** — triage, sorting, formatting, the work measured in throughput — run happily on the lighter end: GPT-5.6 Luna keeps the full 1M context without the heavyweight temperament.
- **The middle** is Terra, Kimi K3, and GLM 5.2 territory: capable defaults for [teammates with a broad remit](/blog/hire-your-second-third-and-tenth-teammate).

There's no wrong first pick, because switching is free and the teammate's memory survives it. Try a lighter brain on a lane for a week; [the record of its work is all in the workspace](/blog/a-company-that-remembers), so you can judge the output, not the vibes — and move it up or down a weight class accordingly.

One team. Different minds. Your call, per teammate — full details in [the changelog](/changelog).