If GPT-5.5 was a very good colleague, GPT-5.6 is a small department with a traffic controller. Sol handles the hard, expensive thinking. Terra takes the normal work. Luna clears the queue. The important launch is not only those three names; it is the place OpenAI put them.
ChatGPT Work is now the agent for jobs too large to fit in a polite little chat bubble. Give it a project, source files, and connected apps; it can gather context, break work into steps, keep going in the background, then produce editable docs, slides, sheets, Sites, or web apps. Scheduled Tasks can keep repetitive work moving. You remain the editor, approver, and person who has to explain the spreadsheet at Monday’s meeting.
On desktop, the former Codex app is merging into the new ChatGPT desktop app. Codex is still the coding agent—now alongside Chat and Work—with inline diff editing, pull-request review, faster computer use, and multi-repository projects. The app also has a built-in browser, while Computer Use can click, type, and move files across your local apps when you approve the access. That is a much bigger deal than a fresh model badge: OpenAI is joining chat, work, code, and desktop actions into one surface.
Pick the brain, not a religion
| Job | Route it to | Why |
|---|---|---|
| Complex investigation, serious codebase work, research with several moving parts | Sol | Highest capability and best agentic-coding results in the family |
| Analysis, writing, normal product work, ordinary multi-step agents | Terra | GPT-5.5-competitive work at half Sol’s API sticker price |
| Classification, extraction, drafting, and high-volume routine automation | Luna | Fastest and cheapest GPT-5.6 tier |
| One giant, parallelizable mission | Sol + ultra | Multiple agents work in parallel; use only when the extra work is worth it |
The reason that routing matters is efficiency. OpenAI reports Sol at 53.6% on Agents’ Last Exam, 80 on the Artificial Analysis Coding Agent Index with max reasoning, and 62.6% on OSWorld 2.0. Those are impressive launch numbers, and the company also reports sharply lower time, tokens, or estimated cost than several competitors on selected tests. They are not magic beans. Benchmark setup, effort setting, tool harness, and the job itself all matter.
The headline usability change is availability: GPT-5.6 launched across ChatGPT, Codex, and the API. But the three-tier family is not a buffet with identical access. Chat users on Plus and above can use Sol at medium or higher effort; Work and Codex expose different plan-based choices; ultra is more restricted. Start by checking the model and effort controls you actually see, rather than buying a plan for a feature that has not reached your account yet.
The honest catch
More agency means more responsibility. A long Work task may consume far more included usage than a chat reply. Connected apps and Computer Use are useful precisely because they can reach real company context, which makes permissions and approval points non-negotiable. And Sol’s stronger safeguards can make legitimate security-adjacent work slower or fussier.
So the winning habit is not “always Sol.” It is Sol for the mountain, Terra for the commute, Luna for the conveyor belt—with ChatGPT Work as the project desk that keeps the three from stepping on each other.
