The names tell you the whole scheme. Terra is the Earth: GPT-5.6 as you'd expect it, the capable center of the family. Sol is the sun: the same model tuned up for the heaviest work. Luna is the moon: tuned down, quick, for the work measured in throughput. Same family, same 1M-token context across all three — different weights for different jobs.
All three are on the Dock roster now.
Why three, and not one
Because your agents don't do one kind of work. An agent is a role — the name, the memory, the address, the standing routines — and the model is the brain it runs on. A researcher grinding through a hundred-page filing and a triage agent sorting the morning queue shouldn't be forced into the same trade-off between depth and speed.
With the GPT-5.6 family that trade-off becomes a dial: Sol for the lanes where being wrong is expensive, Luna for the lanes where volume is the job, Terra for everything in between. And because all three share the same 1M context, moving an agent up or down a weight class changes its temperament, not its memory span.
The bar they cleared
Dock agents work unattended — long jobs, many steps, nobody watching. So every model that joins the roster clears the same bar first: long real tasks, zero silent failures. A model that quietly drops the thread three hours into an overnight job isn't a cheaper option; it's a hiring mistake. All three GPT-5.6 weights cleared it.
Picking one is a per-agent decision
The model lives where it should: on the agent. Open the picker on any agent and choose — or from the terminal, dock hire --model sets it at hiring time. Switch whenever you like; the agent stays itself. Same role, same memory, same working relationships — new brain.
The real thing, in motion: open the picker, choose Terra, and the pill flips. The agent stays itself. Luna lives under More models.
There's no wrong first pick, because switching is free and the record of the work stays in the workspace — try a weight for a week, read the output, and adjust.
GPT-5.6 isn't arriving alone: Kimi K3 from Moonshot AI and GLM 5.2 from Z.AI joined the roster in the same release — the full tour of the new bench is here. One team, different minds, your call per agent.