Ask ChatGPT how to time a Sunday roast and the answer can now arrive as a plan, with the timings laid out beside the recipe. Ask it to split a dinner bill and it can build a small bill splitter right in the conversation. Ask about the Monty Hall problem, the famously counterintuitive game-show puzzle, and you can play it out yourself.
This is Intelligent UI, which arrived with GPT-6 in ChatGPT on October 7. OpenAI trained the model to compose answers from text, visuals and interactive pieces such as buttons, forms and charts, picking whichever format fits the question, plain text included. Underneath is a library of native components and a compiler that renders the interface progressively, so it appears while the model is still writing. OpenAI sums up the ambition in one line: "Instead of people adapting to software, software will adapt to people."
It is a long way from where ChatGPT began. OpenAI launched it on November 30, 2022, as a conversational model that could handle follow-up questions and own up to its errors. The company itself was introduced in December 2015 as a nonprofit research lab. Today the nonprofit OpenAI Foundation governs OpenAI Group, a public benefit corporation, and the stated mission is to ensure that artificial general intelligence, which OpenAI describes as AI systems generally smarter than humans, benefits all of humanity. ChatGPT now reaches more than 1.2 billion people every week.
A lineup with celestial names#
The GPT-6 family reads a little like an astronomy lesson. GPT-6 Astra, introduced in September, is OpenAI's most capable model, aimed at the most demanding professional work: software engineering, research, cybersecurity and computer use, in which the model operates software through the screen, clicking and typing the way a person does. OpenAI says Astra can work through everyday business applications even when they have no API, the programmatic doorway that lets one piece of software talk to another.
GPT-6.1 Sol offers close to Astra's performance at a lower cost for complex coding, research and computer use. GPT-6.1 Terra, released October 26, is the fast, efficient member, good at reading visual material, precise computer use and clear writing. OpenAI's examples run from drafting proposals and interpreting technical drawings to filling out forms and editing video. GPT-6 Luna is the most efficient of all, built for focused, high-volume jobs like pulling fields out of invoices or sorting incoming requests.
The flagship models share a context window, the amount of material a model can consider at once, of about 1.05 million tokens, the word fragments models read. Developers can also dial reasoning effort, roughly how long a model thinks before it answers, from the lowest setting up to max.
The specialists#
GPT-Rosalind is a life-sciences reasoning model for approved organizations, and GPT-5.6 Cyber supports authorized vulnerability research and security testing.
ChatGPT Images 2.5, and the GPT-Image-2.5 Flare and Sunburst models in the API, handle image generation and precise editing. Across ChatGPT and the API, people create more than 3 billion images a week, and a new Sketch feature lets you draw a rough reference by hand.
For voice there is GPT-Live, built on a full-duplex design, meaning it listens and speaks at the same time. It can offer small acknowledgments while you talk, stay quiet while you think, and hand harder questions to a frontier model behind the scenes while the conversation keeps flowing. It powers ChatGPT Voice and is available to developers as GPT-Live-1.
ChatGPT learned to finish the job.#
The bigger change this year is that ChatGPT does the work instead of only describing it. ChatGPT Work, launched in July, is an agent: a model that plans its own steps and uses tools to carry them out. Give it a goal and it gathers information across your apps and files, then turns it into finished sheets, slides, docs and web apps, staying with a project for hours by breaking it into smaller steps. You can follow along, answer its questions, change direction and approve important actions, and Scheduled Tasks keep projects moving while you are away.
Plugins connect it to Slack, Microsoft Teams, Google Drive, SharePoint, email, calendars and CRMs. The ChatGPT desktop app for Mac and Windows goes further, working with local files and apps and a built-in browser. Sites, in public beta, turns work into an interactive site or web app you can share by link: a live dashboard, a launch calendar, a quick prototype.
Then there are dots, introduced in late September. Dots are always-on agents powered by GPT-6 Astra, each with its own cloud computer and browser. A dot can connect to more than 4,000 apps through plugins, answers in ChatGPT, Slack or Teams, learns your preferences from feedback and works toward your goals around the clock. OpenAI describes an early tester whose dot noticed a forgotten invoice, prepared it and sent it once he approved. Dots are rolling out on Pro, Business Premium and Enterprise plans in eligible markets.
Codex, the coding agent that outgrew code#
Codex began as OpenAI's coding agent for developers, and it still is one. It writes and reviews code, works across several repositories in one project, and lets you edit inline within a diff, the side-by-side view of what changed.
More than 5 million people use it every week, and about a fifth of them are not developers at all. Analysts, marketers, designers and bankers use role-specific plugins that bundle apps, skills and workflows for data analysis, creative production, sales, product design and investing.
In July the Codex app merged into the new ChatGPT desktop app, and Codex technology now sits inside ChatGPT Work. Developers who prefer other surfaces still have the Codex CLI for the terminal, Codex Cloud, and an SDK for building Codex into their own tools.
Developers get the same parts.#
The OpenAI API is the interface that lets developers call these models, all listed in the model catalog, from their own software. The Responses API pairs model calls with built-in tools such as web search, file search, a code interpreter and computer use.
The Agents API, in public beta since September, opens up the same harness and infrastructure that run Codex. A developer can create an agent in a single call by naming the task, model, tools and environment, then run it in an OpenAI-hosted sandbox, on their own infrastructure or with a sandbox partner. It compacts older context automatically as a session runs long, loads tool definitions only when they are needed, and supports the Model Context Protocol, an open standard for connecting agents to outside tools. The Agents SDK serves teams who want to orchestrate agents in their own code.
Developers can also build plugins that add skills, tool servers and optional interface elements to ChatGPT and Codex, let people sign in with ChatGPT, and build voice apps on the Realtime API and GPT-Live.
For teams that want a model running entirely on their own hardware, OpenAI's gpt-oss models, released in August 2025 under the Apache 2.0 license, are open-weight: the trained model itself can be downloaded. The larger one runs on a single 80 GB GPU, and the smaller one on devices with 16 GB of memory.
Beyond OpenAI's own platform, Microsoft remains its primary cloud partner, its frontier models and Codex are on AWS, and a Google Cloud integration is due in the coming weeks.
Research at the scale of millions of cells#
GPT-Rosalind offers a glimpse of where this is heading in science. In a project published in October, OpenAI, PerturbAI and Scripps Research set thousands of Rosalind-powered agents to work on existing data from millions of mouse brain cells: a whole-brain atlas of what happens when individual genes are disrupted.
Each agent took one gene disruption, analyzed how the cells responded, searched the scientific literature and proposed an explanation. The run took 32 minutes, where the team estimated a comparable human review would take months to a year. The agents reproduced known findings and surfaced new leads, including a possible link between cholesterol regulation and neurodegeneration, and their judgments of novelty lined up with those of blinded human experts.
The resulting reasoning atlas is open for researchers to explore, and a new academic plan gives eligible labs sponsored access to GPT-Rosalind.
Safety, written down#
OpenAI publishes the frameworks it holds its models to. The Preparedness Framework tracks risks from frontier capabilities and the safeguards required as those risks rise. Astra is the first OpenAI model to reach the framework's Critical threshold for cybersecurity capability, and it shipped with strengthened protections against misuse and against the model taking unauthorized actions.
The Model Spec, first published in 2024, is OpenAI's public description of how its models should follow instructions, resolve conflicts and behave safely, written so that users, researchers and policymakers can read and argue with it. Major models arrive with system cards that document their safety testing.
In ChatGPT Work and Codex, businesses can also set confirmation policies that require a person's approval before consequential actions. And audio generated with GPT-Live carries SynthID watermarking, with a public tool and an API for checking where a clip came from.
Why it matters#
From a chat window that designs its own interface to agents with their own computers, OpenAI offers almost every layer of an AI product ready to use.
That breadth puts the same building blocks in the hands of a student with a homework question, a startup shipping its first agent and an enterprise rebuilding its support desk. The oldest advice still applies: open ChatGPT and ask for something a little ambitious.
Ausprobieren