OpenAI Unveils GPT-5.6 and ChatGPT Work: Three-Tier Model Goes Live, AI Starts "Taking Over" Your Work
After being delayed by the US government for a week, GPT-5.6 officially launched on July 9: OpenAI's Sol, Terra, and Luna three-tier models, along with ChatGPT Work, which can work continuously for hours, are moving the battlefield from chat windows to offices.
On the evening of July 9, a marketing consultant in Taipei opened ChatGPT as usual to write her weekly report. She discovered that the model menu had changed: the familiar model names were gone, replaced by three new names - Sol, Terra, and Luna. She casually handed over a project requirement to the new ChatGPT Work and went to take a shower. When she came back, a draft report, budget table, and a one-page website prototype were already on the screen.
This is what OpenAI wants you to experience: AI is no longer just "you ask, it answers," but "you assign tasks, it delivers results."
Event Background
GPT-5.6 was originally scheduled to be released at the end of June, and OpenAI had already announced the names of the three new models. However, the release was put on hold by the US government due to national security concerns - a move consistent with the brief export control of Claude Fable 5 in June, indicating that Washington is starting to have a say in the release of top models (see our report on Fable 5's restoration of export controls). It wasn't until July 9 that OpenAI obtained permission to officially release GPT-5.6.
The naming logic has also changed: the number represents the "generation," and Sol, Terra, and Luna are three "capability layers" that can exist for a long time - Sol is the flagship that handles the most difficult tasks, Terra is the daily workhorse, and Luna is fast and cheap. In the future, each capability layer can be upgraded at its own pace without waiting for the entire generation to be updated.
Key Points
- GPT-5.6 three-tier model fully launched: Sol (flagship), Terra (balanced), and Luna (fast and cheap), with API pricing of $5/2.5/1 per million tokens for input and $30/15/6 per million tokens for output.
- ChatGPT Work agent debuts: can read the context of your applications and files, break down large projects into small tasks, produce reports, spreadsheets, documents, and even web apps, and execute continuously for several hours.
- Codex merged into ChatGPT desktop version: the ability to write code is no longer a separate product, and the desktop app is now a one-stop shop.
- New website hosting service added: AI-generated web pages can be directly hosted by OpenAI.
Market Impact Analysis
For Taiwanese users: ChatGPT Work is aimed at "office workers' entire projects" - proposals, reports, and presentations. In the past, you might have used ChatGPT to write an outline and then manually create a PowerPoint presentation. Now, the agent can directly deliver the finished product. Honestly, this is a productivity bonus for "those who can give instructions," but a warning for "those who only know how to make presentations."
For enterprise applications: this move directly encroaches on Anthropic's Claude Cowork and Microsoft Copilot's territory. Companies need to start thinking about which processes are worth handing over to agents for comprehensive processing. Teams that want to get started can begin with small scenarios in automated tasks.
For developers: the three-tier pricing makes it clearer to distinguish between "cheap and large-scale" and "expensive but smart." Luna's input price of $1 per million tokens will make many small and medium-sized applications recalculate their cost structure; after Codex is integrated into the desktop version, the agent development war with Claude Code tools will only become more intense.
Future Development Trends
The naming system of "number = generation, name = capability layer" reveals OpenAI's product rhythm: future upgrades will be like regular, layered mobile system updates. More noteworthy is the government's review and intervention in the release schedule - the release of top models now has an additional "Washington approval" step. The model competition is shifting from "who is the smartest" to "who can be the smartest while complying with regulations."
TheAI Academy Summary and Comments
Comments: GPT-5.6 is an incremental upgrade, but the real focus is on ChatGPT Work - OpenAI is officially commercializing "AI colleagues," and the era of chatbots is beginning to come to an end.
Practical suggestions for Taiwanese readers: 1. Try ChatGPT Work this week and hand over a task that you would normally spend half a day on (such as organizing a price list and creating a presentation) to experience the difference between agents and chatbots; 2. Developers should use Luna to calculate the cost of running their existing applications and save money where possible; 3. Don't rush to cancel other tools - the quality of agent-delivered products still requires human supervision, and your professional judgment is more valuable now.
Data Sources
- Axios: OpenAI releases GPT-5.6 and ChatGPT Work tool
- Engadget: OpenAI gets permission to roll out GPT-5.6 on July 9
- The Hill: OpenAI announces GPT-5.6 release after delay
(This article is compiled based on publicly available information and is subject to official announcements.)
Choosing the Right Capability Layer
In GPT-5.6's three capability layers, Sol, Terra, and Luna each have their pros and cons. The following table compares the characteristics of these three capability layers:
| Capability Layer | Processing Ability | Pricing (Input) | Pricing (Output) | Suitable Use |
|---|---|---|---|---|
| Sol | Flagship-level | $5/1 million tokens | $30/1 million tokens | Tasks that require high precision and complex processing |
| Terra | Balanced-level | $2.5/1 million tokens | $15/1 million tokens | Daily work and general tasks |
| Luna | Fast and cheap | $1/1 million tokens | $6/1 million tokens | Tasks that require fast processing and low cost |
It is crucial to choose the right capability layer based on your needs and budget. For example, if you need to process large amounts of text data, Luna might be the most suitable choice. However, if you need high precision and complex processing, Sol might be a better choice.
Common Misconceptions / Debunking Myths
There is a common misconception that GPT-5.6's three capability layers are ranked in order, with Sol being the best and Luna being the worst. However, in reality, these three capability layers are designed for different needs and uses. Sol is not necessarily the best choice, and Luna is not the worst choice. What matters is choosing the right capability layer based on your needs and budget.
Practical Steps
To start using GPT-5.6 and ChatGPT Work, follow these practical steps:
- Log in to ChatGPT Work and choose the right capability layer.
- Upload your text data or task requirements.
- Set up the input and output parameters.
- Start ChatGPT Work and wait for the results.
- Check the results and make modifications as needed.
Future Trends
The release of GPT-5.6 and ChatGPT Work marks a new era for AI technology. In the future, AI will be increasingly applied to various fields, including business, education, and healthcare. As AI technology continues to evolve, ChatGPT Work and GPT-5.6 will be continuously upgraded and improved, providing better services and features.
Suggestions for Different Groups
- For enterprise users, GPT-5.6 and ChatGPT Work can help improve productivity and efficiency, saving time and cost.
- For developers, GPT-5.6 and ChatGPT Work provide powerful APIs and tools to help develop better AI applications.
- For individual users, GPT-5.6 and ChatGPT Work can help complete various tasks, including text editing, data analysis, and creative writing.
Frequently Asked Questions
What are the differences between GPT-5.6's Sol, Terra, and Luna models?
Sol is the flagship model for handling the most challenging tasks, Terra is the daily balanced mainstay, and Luna is the fastest and most affordable option. API pricing is $5, $2.5, and $1 per million tokens input, and $30, $15, and $6 output, with the numbers representing generations and names representing capability layers.
What is ChatGPT Work?
ChatGPT Work is a work agent launched by OpenAI on July 9, 2026, alongside GPT-5.6: it can read application and file contexts, break down projects into smaller tasks, produce presentations, spreadsheets, documents, and web apps, and execute continuously for hours, with the desktop version integrating Codex, a programming agent.
Why was GPT-5.6 delayed?
OpenAI originally planned to release GPT-5.6 at the end of June, but the US government requested a delay for review, citing national security concerns, and it was finally approved for public release on July 9. This is the second time Washington has intervened in the release schedule of a top AI model, following the export control incident involving Claude Fable 5.
Do general users need to upgrade to GPT-5.6 immediately?
The difference is limited for daily Q&A, but if your work involves delivering complete projects, such as presentations, reports, and documents, ChatGPT Work is worth trying first; developers are advised to evaluate whether Luna's low pricing can reduce the cost of existing applications.