# What are GPT-5.6 Sol, Terra and Luna, and what is ChatGPT Work?

Date: 2026-07-10 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-are-gpt-5-6-and-chatgpt-work/

**Answer:** GPT-5.6 is OpenAI's new family of models, released for everyone on July 9, 2026, in three tiers: Sol, the most capable, Terra, a balanced everyday model, and Luna, the cheapest and fastest. On the same day OpenAI launched ChatGPT Work, an agent inside ChatGPT that pulls information from your connected apps and keeps working for hours to deliver finished documents, slides, spreadsheets and small web apps.

Yesterday, Thursday, July 9, 2026, OpenAI moved its GPT-5.6 models out of a limited preview and made them available to the public in ChatGPT, in Codex, its coding agent, and in the API that developers use to build their own apps. The release comes with a new naming idea: instead of one model with a "mini" sibling, GPT-5.6 is a family of three tiers named **Sol**, **Terra** and **Luna**.

Alongside the models, OpenAI introduced **ChatGPT Work**, an agent in ChatGPT built on Codex technology. An **agent** is an AI system that pursues a goal over many steps, choosing tools as it goes, rather than giving one reply. It is aimed at office tasks: budgets, briefs, reports and presentations.

## The three tiers, explained

OpenAI says the number in the name marks the generation; the three names are lasting tiers that can each be improved on their own schedule.

| Tier | Role | API price per million tokens (input / output) |
|---|---|---|
| GPT-5.6 Sol | Flagship, for the hardest work | $5 / $30 |
| GPT-5.6 Terra | Balanced, for everyday tasks | $2.50 / $15 |
| GPT-5.6 Luna | Fastest and most affordable | $1 / $6 |

A **token** is a small chunk of text, often part of a word, and API prices are counted per million of them. Input tokens are what you send; output tokens are what the model writes back.

OpenAI's theme for this release is efficiency: getting more finished work out of each token. Its own figures say Sol scores 80 on the Artificial Analysis Coding Agent Index, an independent ranking of coding agents, slightly ahead of Anthropic's Claude Fable 5, while using fewer than half as many output tokens. TechCrunch reports that Sam Altman told CNBC Sol is 54% more efficient with tokens when coding. These are company claims; independent testers will take time to check them.

GPT-5.6 also adds two higher **effort settings**, which control how long the model thinks before answering. **max** gives it more time to reason, test ideas and revise. **ultra** goes further and runs four agents in parallel by default, each working on part of a problem, then combines their results.

For developers, a feature called **Programmatic Tool Calling** lets the model write and run a small program that calls tools and filters their results, so it does not have to read every intermediate result itself. That saves tokens on tool-heavy tasks.

## What ChatGPT Work does

ChatGPT Work takes a goal, such as "prepare for Monday's sales meeting," and turns it into a deliverable. It gathers information through **plugins**, connections to apps like Slack, Microsoft Teams, Google Drive, SharePoint, email and calendars. You can point it at one app by typing @ followed by the app's name. It then breaks the project into smaller steps and works through them, sometimes for hours.

Several pieces come with it:

- **Scheduled Tasks** run a job once, on a schedule, when something happens, or to watch for changes, for example refreshing a meeting agenda from Slack every week.
- **Sites in ChatGPT**, in public beta, turns work into an interactive web page or small web app you can share by link, such as a project tracker or a live dashboard.
- On the **desktop app**, ChatGPT Work gets a built-in browser and **Computer Use**, meaning it can click, type and move files in your own apps. The Codex app is merging into this new ChatGPT desktop app, and OpenAI says it will begin winding down its standalone Atlas browser.

You stay in charge: you choose what it can access, when it checks in and which actions need your approval. For companies, an **auto-review** step uses OpenAI's strongest models to inspect important actions in connected tools before they happen, to catch unauthorized sharing of sensitive data.

## Who can use it today

- **ChatGPT chat:** Plus, Pro, Business and Enterprise users get Sol; Pro and Enterprise can also pick Sol Pro.
- **ChatGPT Work and Codex:** people on the Free and Go plans get Terra; paid plans can choose any tier and an effort level.
- **ChatGPT Work on web and mobile:** rolling out first to Pro, Enterprise and Edu, then to Plus and Business over the next few days.
- **Desktop app:** Chat, Work and Codex on every plan, including Free, for Mac and Windows.
- **API:** all three tiers, rolling out over 24 hours.

## A worked example

OpenAI suggests starting ChatGPT Work with a task you already know well, so you can judge the result. A clear request names the sources, the output and the point where it must stop:

```prompt
Using the "Q2 budget" spreadsheet in my Google Drive and last month's
finance channel in Slack, explain the three largest differences between
planned and actual spending. Put the result in a two-page document with
one summary table. Do not share or send anything; show me the draft first.
```

The first sentence tells it which plugins to use, the second defines the finished product, and the last keeps a person in control of anything that leaves your account.

> **Try it:** Pick a weekly report you already write by hand. Describe it to ChatGPT Work as a request with its sources, format and deadline, then compare its draft with your last real one line by line.

## Why it matters

GPT-5.6 is OpenAI's clearest bet so far that the next step for AI is finished work, not better chat answers. TechCrunch reads the launch as aimed mainly at Anthropic, whose Claude models have won many business customers, and OpenAI's charts compare Sol directly with Claude Fable 5 and Opus 4.8.

Security is the other thread. OpenAI calls GPT-5.6 its strongest cybersecurity model yet, but says it stays below the "Critical" level of its Preparedness Framework, the internal scale it uses to rate dangerous abilities. TechCrunch notes that the US administration had earlier tried to restrict the rollout over misuse fears. OpenAI's answer is layered safeguards that it says block about ten times more potentially harmful activity, plus an option in ChatGPT and Codex to retry a blocked request on a less capable model when the block is a mistake.

## Which GPT-5.6 tier should you pick, and how do you start with ChatGPT Work?

For most everyday questions, writing and analysis, **Terra** is a sensible default: it is what free users get in ChatGPT Work and Codex, and OpenAI says it outperforms the earlier GPT-5.5 at lower cost. Choose **Sol** when the task is long, technical or high-stakes, and **Luna** when speed and price matter most, such as sorting many short texts through the API. Keep max and ultra for problems where extra time clearly pays off.

To start with ChatGPT Work, connect only the plugins one task needs, give it a goal you can check, and ask for a draft before anything is shared. Watch its progress the first few times and redirect it when it drifts. If you schedule recurring jobs, review the first few results before trusting them to run on their own.

## Sources

- GPT-5.6: Frontier intelligence that scales with your ambition: https://openai.com/index/gpt-5-6/
- ChatGPT is now a partner for your most ambitious work: https://openai.com/index/chatgpt-for-your-most-ambitious-work/
- OpenAI launches its new family of models with GPT-5.6: https://techcrunch.com/2026/07/09/openai-launches-its-new-family-of-models-with-gpt-5-6/