OpenAI turns the agent wheel: GPT-5.6, Work and Atlas phased out
OpenAI consolidates agent features in ChatGPT Work, opens up GPT-5.6, and sunsets Atlas. Plus: research on self-improvement and an open-source workflow.
Inhaltsverzeichnis
Today, almost everything revolves around the same question: How much work should AI take over on its own? OpenAI offers several answers at once — with a new work mode, a more open model, and the next browser cull. At the same time, research and open source show where agentic workflows are headed.
🤖 OpenAI launches ChatGPT Work and opens up GPT-5.6
OpenAI is launching ChatGPT Work, an agentic work environment based on Codex and the now freely available GPT-5.6. The idea: you define a goal, and the agent then works through tools such as Google Drive, Slack, or Salesforce and handles multi-step tasks largely independently. Depending on the plan, it runs on web, mobile, and desktop. Source
This is more than just a new feature label. With this, OpenAI is clearly shifting ChatGPT from an “answer machine” to a “work surface.” For knowledge work, that sounds attractive: less context switching, less copy-paste, more automation. At the same time, the requirements for permissions, auditability, and error control are increasing. If an agent is rummaging around in Slack or Salesforce, you want to know very precisely what it may do — and what it may not. The classic AI moment of “I almost got it right” gets expensive fast in an enterprise context.
🧭 Atlas is discontinued, browser functions move into ChatGPT
OpenAI is discontinuing its AI browser Atlas after less than eight months and shifting the most important functions directly into ChatGPT. Instead of a separate browser, the company will in future rely more heavily on Computer Use and browser-like actions within the ChatGPT app and via a Chrome extension. Source
This is a classic OpenAI move: pack a product back up, redistribute the features, move on. For users, that is ambivalent. On the one hand, a standalone tool disappears; on the other, usage becomes simpler because you don’t have to learn yet another interface. Strategically, though, the message is clear: the browser as an independent AI product apparently isn’t strong enough yet, while ChatGPT as the central control hub keeps accumulating more and more tasks. Anyone who bet on AI browsers as a major category is probably staring at their coffee a bit longer today.
🧪 GPT-5.6 Sol is said to have trained a smaller model itself
According to OpenAI, GPT-5.6 Sol post-trained a smaller model called Luna using a single, fairly vaguely formulated prompt. In the internal RSI benchmark for recursive self-improvement, Sol is said to be 16.2 points ahead of GPT-5.5. OpenAI is already talking about the “automated researcher” being within reach. Source
Why does this matter? Because here it’s not just one model answering, but one model supporting other models in learning. That is an important building block toward AI Research Automation: AI that experiments, evaluates, and improves. The catch remains the same as always with self-improvement: benchmarks are not reality, and an internal result is not yet robust general capability. Still, the direction is exciting — and probably much more important for model development than the next pretty demo.
⚖️ OpenAI under pressure in copyright dispute
In the ongoing copyright case, the New York Times accuses OpenAI of withholding evidence. At its core, the dispute is about how training data, documentation, and possible disclosure obligations were handled. Such proceedings are long since no longer a side note; they are shaping the rules of the game for the entire AI industry. Source
For you as a user or company, this is not just legal background noise. Copyright and disclosure issues affect which models are available, how they are trained, and how providers secure their products. The larger the systems and the broader their use, the more important data provenance, licenses, and traceability become. The big AI providers are not only building products right now, but also defense lines — often in the same week.
🛰️ LiST: More robust and better-calibrated networks
With LiST – Lipschitz Scaling Training for Robust and Calibrated Neural Networks, there is a new approach on arXiv that addresses robustness and calibration together. The core idea: instead of treating these goals separately, LiST tries to train models so that they respond more reliably to inputs and estimate their uncertainty better. This is particularly relevant for applications where wrong decisions are costly. Source
That may first sound like classic research work, but practically it is quite important. A model that tells nonsense “with high confidence” is no better than an uncertain model — just more self-assured. Especially for foundation models in sensitive areas such as Earth observation, medicine, or industry, it is not only accuracy that matters, but also calibration. Research is slowly moving away from mere “it’ll probably be fine” toward “please be reliable.” The only revolutionary thing about that is how long it took.
🛠️ Tool tip of the day: Model_Forge for agentic 3D workflows
Model_Forge is an open-source prototype for AI-assisted 3D printing workflows. The project orchestrates steps such as CAD generation, validation, a human-approval gate, slicing, and artifact tracking. For anyone who doesn’t just want to discuss agentic pipelines but wants to build them in practice, this is an exciting toolkit. Source #
What is especially interesting here is the workflow mindset: not a single agent does everything, but multiple steps are connected in a controlled way. That is exactly what many productive agent systems look like in practice. You get less magic, but significantly more control — and in real projects, that is usually the better news. If you are experimenting with agentic AI yourself, it is worth taking a look at prototypes like this as a blueprint for clean handoffs, checks, and approvals. #
Don’t want to miss any news? Subscribe to the newsletter