ChatGPT Work is officially launched, powered by Codex + GPT-5.6, and integrates directly with Slack, Teams, Google Drive, SharePoint, CRM, and more to execute complete workflows.

Author and source: 0x9999in1, ME News



TL;DR

• OpenAI launches the GPT-5.6 series with three pricing tiers: Sol (Premium), Terra (Balanced), and Luna (Budget), all simultaneously available on ChatGPT, Codex, and API.
• Sol scored 53.6 in the Agents' Last Exam, outperforming Claude Fable 5 by 13.1 points at approximately one-fourth the cost.
• The Programming Agent scored 80 on the Sol index, 2.8 points higher than Fable 5, but used less than half the tokens and took less than half the time.
• Added Ultra mode, defaulting to 4 parallel Agents; API users can also access the multi-Agent beta.
• ChatGPT Work is officially launched, powered by Codex + GPT-5.6, and integrates directly with Slack, Teams, Google Drive, SharePoint, CRM, and more to execute complete workflows.
• The Codex desktop app has been integrated into the new ChatGPT desktop app, combining Chat, Work, Codex, and Browser into one unified platform.
• Atlas Browser will shut down on August 9, less than a year after its release, becoming OpenAI's first explicitly discontinued standalone product.

I. This time, OpenAI isn't just upgrading—it's reorganizing its formation.

First, the conclusion.

The GPT-5.6 series is not a routine update. It represents OpenAI’s strategic repositioning in the second half of 2026.

Model layer: Fully covered with Sol, Terra, and Luna tiers. Product layer: ChatGPT Work is now elevated. Desktop layer: Codex app integrated into ChatGPT; Atlas discontinued.

Happened on the same day. A lot of information.

Viewed separately, these three items are three news stories. Viewed together, they form a strategy.

OpenAI is telling the market: I want models, agents, and entry points—all of them—but I’m no longer letting them operate in isolation.

Two, Sol's numbers are impressive, but it depends on how you use them.

First, look at the model.

GPT-5.6 Sol scored 53.6 on the Agents' Last Exam, which is 13.1 points higher than Claude Fable 5. These are the official figures.

What does 13.1 points mean?

In agent evaluations, a single-digit lead is considered significant. A double-digit lead essentially means they’re not in the same league. Even when reducing reasoning intensity to medium, Sol still leads by 11.4 points, with estimated costs roughly one-quarter of Fable 5’s.

One quarter. These four characters are the key.

In the past, people were accustomed to a default logic: stronger models = more expensive. If you want accuracy, you pay. If you want affordability, you accept the trade-off.

Sol broke this default by simultaneously pushing performance and price in two opposite directions.

The same applies to the programming side: Artificial Analysis’s programming agent score for Sol is 80, and for Fable 5 it’s 77.2—the difference is not significant, just 2.8 points.

But what really hurts is the process data:

• Output token, Sol is less than half of Fable 5.
• It took less than half the time to complete.
• Cost reduced by approximately one-third.

The difference in spread is small, but the efficiency is overwhelming.

This means Sol doesn’t achieve its score by overthinking. It takes shorter, more accurate, and more efficient reasoning paths.

For businesses, this is the most straightforward accounting: for the same task, a model that takes less time, uses fewer tokens, and charges less is what the CFO will notice.

III. Ultra Mode: Bundle Parallel Computing Power into a Single Button

GPT-5.6 also introduces the Ultra mode.

It defaults to running 4 agents in parallel. Use more tokens for stronger results and shorter processing times. API users can also access the multi-agent beta.

This design isn't new. Methods such as multi-agent systems, voting, and self-consistency have been around in papers for many years.

The new one is productized.

Previously, if you wanted four models to work on a problem together and then select the best answer, you had to write the orchestration yourself, bear the costs yourself, and fine-tune the prompts yourself.

Now, OpenAI has turned it into a switch.

Regular users can click on Ultra to access reasoning strategies previously available only to researchers and senior engineers.

The significance of this step lies not in the algorithm, but in the barrier.

It transforms "brute-force parallelism" from an engineering technique into a consumer-grade option.

Four: Price Table: A Naked Price War

Looking at the API pricing together, it's very straightforward.

Per 1 million tokens:

• Sol: Input $5, output $30.
• Terra: Input $2.50, output $15.
• Luna: Input 1 USD, output 6 USD.

What does this price mean?

The input price for flagship Sol is $5 per million tokens. At this rate in 2024, you could only afford a mid-tier model; by 2026, it will be considered flagship.

Terra Balanced Edition, input $2.5. Ideal for most internal enterprise applications.

The budget version of Luna: input $1, output $6. This price is already competitive enough to match all open-source deployment solutions.

OpenAI didn't use opaque pricing this time. Three clear tiers with distinct gradients and well-defined features.

People with decision fatigue will feel much more comfortable.

For Anthropic, Google, and a host of open-source models, Luna’s purpose matters more than Sol—because Sol is designed for clients with unlimited budgets, while Luna is for those who want to save money without compromising.

This is the real level where the price is being suppressed.

Five: ChatGPT Work: An Agent That Takes Over Your Desk

After the model finishes speaking, review the product.

ChatGPT Work is the most significant feature in this release.

It's not chatting. It's taking on work.

Give it a goal, and it will connect to your Slack, Teams, Google Drive, SharePoint, email, calendar, CRM, and project management tools. It reads the context and generates documents, spreadsheets, presentations, analytical reports, and even web applications.

Users can @ an app to have it extract information from it.

What is this?

This is the first time OpenAI has clearly declared: ChatGPT is no longer just a "knowledge-based conversational tool." It aims to be a "work execution system."

Over the past two years, the term "Agent" has been overhyped. Since AutoGPT, countless projects claiming to "automatically get things done" have emerged, but most don’t last more than three months.

The reason is simple: being able to converse doesn't mean you can get things done; being able to get things done doesn't mean you can do them correctly; being able to do them correctly doesn't mean you can do them correctly consistently.

ChatGPT Work daring to launch at this time indicates that OpenAI has confidence in the stability of Codex + GPT-5.6.

Whether external parties can replicate this confidence will depend on user feedback after the platform has been open for several weeks.

But the direction is clear: the office software industry is being rewritten.

Notion, Asana, ClickUp, and Monday all derive their core value from "integrating information flows + structured outputs." ChatGPT Work does the exact same thing, just through a conversational interface.

One entry point, replacing ten SaaS platforms. This is the phrase all SaaS vendors fear hearing most by 2026.

Six: Desktop Integration: OpenAI Finally Admits Users Don’t Want Three Separate Apps

Now look at the desktop.

The Codex app will be integrated into the new ChatGPT desktop application, which will offer Chat, Work, Codex, and a built-in browser all in one.

This step should have been taken long ago.

For over a year, OpenAI’s desktop product line has been fragmented: ChatGPT in one app, Codex in another, and Atlas in a browser. Users had to switch between three windows to accomplish a single complex task.

For average users, this isn't "feature-rich"—it's "cognitive overload."

The new desktop app packs four tasks into one window:

• Chat: Conversation and Support
• Cross-app Agent
• Codex: Write code, run code.
• Browser: scrape web pages, open tabs, download files.

This combination is somewhat like a fusion of an IDE, a browser, and an assistant.

And free users can use it too. This is something few people notice, but it’s important.

Free users gaining access to the desktop versions of Chat, Work, and Codex means OpenAI aims to make "using ChatGPT for work" a fundamental habit.

Once a habit is formed, paid conversion is just a matter of time.

Seven: Atlas's Farewell—The First Product Cut, Less Than a Year Later

The most dramatic part is Atlas.

Atlas is a browser independently launched by OpenAI. It was shut down on August 9, less than a year after its release.

This is the first time OpenAI has explicitly abandoned a standalone product that was already launched.

Keep in mind that even features previously "ignored" by OpenAI were rarely officially announced as discontinued. This time, by providing a clear shutdown date, their stance is unmistakable.

Why?

Three reasons, put simply, are all very straightforward.

First, the browser is not the entry point—the Agent is.

Atlas wants to replicate Chrome’s logic: control the browser, and you control users’ online time. But the reality of 2026 is that users increasingly don’t want to browse themselves—they want agents to browse for them.

Browsers are for people to use. Agents are for people to command. These two things naturally have different product forms.

Second, Chrome's moat is deeper than imagined.

Atlas wants to steal users from Chrome, but it’s much harder than imagined. Chrome has already built a complete ecosystem of user migration costs—bookmarks, extensions, password syncing, and account systems. Atlas has nothing comparable.

Even if OpenAI leverages its model capabilities to differentiate itself, it cannot reverse the gravitational pull of the "browser" category.

Third, it's time to consolidate internal resources.

Maintaining three clients—ChatGPT, Codex, and Atlas—at the same time is a burden for any company. Discontinuing Atlas and integrating browsing capabilities as a Chrome extension combined with a built-in browser would save manpower while preserving the functionality.

One notable point in the service suspension notice: Atlas bookmarks, tabs, and history will not be automatically migrated. Users must manually export their data before the service ends.

This statement is simple but harsh. It means OpenAI has no intention of implementing a "smooth transition." Atlas users' assets have, in effect, been "invalidated."

A product less than a year old being handled this way by its own parent company is uncommon in Silicon Valley.

But it also illustrates one thing: OpenAI has moved beyond the phase of trying everything. It is learning to simplify.

Eight: Look at these three things together: This is OpenAI's year of convergence.

From 2024 to 2025, OpenAI was adding on.

Launch new models. Introduce new products. Roll out new applications. Pioneer new forms. GPT-4o, o1, o3, Sora, Canvas, Atlas, Codex, GPT Store… Dozens of releases in a single year.

By 2026, the pace had changed.

The GPT-5.6 series has been released, replacing previous multiple model branches.

ChatGPT Work has arrived, bringing all Agent capabilities together into one product.

Merge desktop components: unify Chat, Work, Codex, and the browser into a single window.

Atlas has been discontinued, and the independent browser direction has been rejected.

Putting these actions together, it’s clear: OpenAI is converging.

What to consolidate? Consolidate entry points, consolidate branding, and consolidate user perception.

A model series (GPT-5.6).

A desktop app (new version of ChatGPT).

An Agent product (Work).

A development tool (Codex).

An API platform.

The structure has become much cleaner.

This consolidation isn't conservatism—it's confidence. You only eliminate redundant products when you're certain you're in the top tier across all these sectors.

If you're still testing, you preserve all possibilities. You let Atlas keep running, let Codex run independently, and ensure every attempt has a chance to survive.

Only when you determine that it's clear who has won and who has lost in these directions should you dare to subtract.

OpenAI is now daring to simplify.

Nine: Competitive Perspective: Anthropic, Google, and the Open Source Camp Must Respond

For Anthropic, Sol's emergence is a hard blow.

Claude Fable 5 is Anthropic's flagship model released in the first half of 2026, positioned as "the strongest reasoning + the strongest coding." Now, Sol is outperforming it on both fronts—and at a lower cost.

Anthropic’s next steps are clear: either match the reasoning efficiency in the next version, or build unmatched depth in vertical use cases. Price competition is not Anthropic’s strength.

For Google, the challenge with the Gemini series is not model scores, but product integration.

Google's advantage is the Workspace ecosystem. Gmail, Docs, Sheets, Slides, and Drive are all under its control. In theory, Gemini + Workspace is the most seamless agent combination.

But now ChatGPT Work can directly integrate with Google Drive, bypassing Google’s advantage of “data is with me” through the back door.

Google needs to accelerate integrating Gemini deeply into every aspect of Workspace, rather than keeping it at the level of "sidebar Q&A."

For the open-source community, Luna's existence is the greatest pressure.

The value proposition of past open-source models was: low cost, controllability, and self-hostability. But with Luna’s pricing at $1 per million input tokens and $6 per million output tokens, the "low cost" advantage has now been directly challenged by closed-source alternatives.

Meta’s Llama series, Mistral, Qwen, and others must reconsider what their true differentiation is: data sovereignty? vertical fine-tuning? cost-effectiveness under extreme scenarios?

There must be a reason for businesses to choose self-deployment over directly using Luna.

Ten: From the User’s Perspective: Short-term is Strengthening, Long-term is Binding

For individual users, the intuitive experience of this update is: the model is smarter, the desktop is more user-friendly, and the Agent can now actually get things done.

Cool or not? Cool.

But should you stay calm? Yes, you should.

As a product becomes increasingly "do-it-all," users' dependence on it deepens at a visibly accelerating rate.

Previously, when you turned on your computer, you would open your browser first, then your IDE, followed by Slack, and finally ChatGPT. These tools existed in parallel.

Now that you've turned on your computer, you only need to open the new version of ChatGPT. Chat, Work, Codex, and the browser are all inside.

This is a victory of efficiency and the consolidation of entry points.

Once your workflow is embedded in an app, the cost of migrating increases significantly. By then, the question of “Should I switch to a different AI?” will no longer be a model selection issue, but an operating system migration challenge.

This is also the result OpenAI wants.

Eleven: Several Issues to Pay Attention To

At this point, several matters have not been clearly addressed by the official sources and warrant continued inquiry from everyone.

First, how are permission boundaries defined when ChatGPT Work operates across applications?

It connects to Slack, CRM, email, and calendars. The data in these systems is among the most sensitive for businesses. How much operational authority does the Agent have? Who is held accountable if something goes wrong? How are audit logs maintained?

Without resolving these issues, enterprise users may try them out but won't dare to fully adopt them.

Second, will Atlas users' assets really be silently "invalidated"?

Bookmarks, history, and extensions require manual export if there is no migration tool. It remains unclear how much dissatisfaction this will cause before August 9.

Third, could the multi-agent billing system get out of control?

The Ultra mode defaults to running 4 agents in parallel. This means token consumption for a single task may be several times higher than in Standard mode. While this has minimal impact on individual subscribers, it can cause noticeable fluctuations in billing for enterprise users charged by API usage.

Fourth, after Codex is merged into ChatGPT, will the pure developer user base be diluted?

Codex originally had its own distinct community character. After merging with ChatGPT, the product language, UI logic, and update pace had to accommodate mainstream users. Will longtime Codex users feel that "the flavor has changed"?

The answers to these questions will only become clear over the coming months.

Twelve, Final Words

If I had to summarize this release in one sentence, I’d say: OpenAI is moving away from casting a wide net and entering the harvesting phase.

The model received a line of 5.6.

The product received a ChatGPT app.

The agent receives a work entry.

Remove the browser entirely and integrate its capabilities into the main product.

This is what a company does when it truly begins to care about "resource efficiency" and "mental occupancy."

Sol’s score was impressive, Work’s ambition was clear, and Atlas’s departure was decisive. Three events happening on the same day created a strong rhythm, indicating that internal coordination is more aligned than outsiders might assume.

Will it ultimately be redeemable?

At the model level, the numbers are right there—run them and you’ll know if they’re real or not.

At the product level, look at retention and renewals over the next few months.

At a strategic level, watch to see whether OpenAI continues to streamline its operations.

Many are betting on OpenAI to win. Those betting against it are still waiting for the next sign of "overextension."

On July 2026, however, OpenAI sent the opposite signal.

It demonstrates that it knows what to do and what to let go of.

This is the hardest part.

Everything else is just a numbers game.

Reference materials

1. OpenAI official blog, *Introducing GPT-5.6: Sol, Terra, and Luna*, July 2026.
2. Official OpenAI announcement, *ChatGPT Work is now available*, July 2026.
3. OpenAI official announcement: *Atlas will be sunset on August 9, 2026*, July 2026.
4. Artificial Analysis, *Coding Agent Index — July 2026 Update*
5. Agents' Last Exam Leaderboard, July 2026 Edition.
6. Anthropic official documentation, *Claude Fable 5 Model Card*, first half of 2026.
7. The Information, *OpenAI's Product Consolidation Play*, July 2026.
8. Bloomberg, *OpenAI Kills Atlas Browser Less Than a Year After Launch*, July 2026.