Sam Altman told TIME that OpenAI will possess a system he defines as artificial general intelligence by the end of 2026.
The company is not there yet, but Altman expects to cross that threshold within two years. Chief Research Officer Mark Chen estimates the company is 80% of the way there. Co-founder Greg Brockman believes observers will soon view this current period as the moment AGI emerged.
These claims come from a report by Alex Heath, who spent more than two weeks interviewing over 20 executives, employees, investors, and rivals.
Astra is meant to work as an automated AI research intern
Central to this confidence is Astra, a model family announced last year. It is designed to function as an automated AI research intern.
Chief Scientist Jakub Pachocki told TIME that Astra already meets this internal benchmark. Hand it an experimental idea and it can write the code in OpenAI’s codebase, run the experiment, and report back with results. Give it a research paper and it can handle work that would have taken a human researcher about a week.
Astra also enables persistent agents, virtual coworkers that tackle tasks on their own over longer periods. Altman sees the biggest payoff in generating new knowledge. “I expect this will be the first model where the model actually invents new things in a way that matters,” he said during a customer preview, according to TIME. “That’s a very AGI-like thing.”
The underlying wager is that a system like this could kick off recursive self-improvement, where AI helps make itself better. Opinions are split. Some researchers think that kind of capability is still far off. Others see early signs it is already happening.
More broadly, there is real debate over whether AI built purely on language models can truly make new discoveries and generalise, which seems like a basic requirement for AGI. The term itself has no precise definition, though, so it is partly a matter of interpretation. OpenAI defines AGI as highly autonomous systems that outperform humans at most economically valuable work, a heavily economics-driven reading. Others argue that more is needed, like a robust understanding of the world, and that LLMs are just one piece of the puzzle.
AI puck hardware expected in early 2027
OpenAI is working on a small handful of devices, according to Altman: something that belongs on a table, something for your pocket, and something you wear. The first will be a small, puck-shaped device that senses its surroundings and talks to its owner via voice mode. It is expected early next year. Altman added that OpenAI will definitely build humanoid robots as well.
Beyond the lofty ambitions, OpenAI has to solve a much more mundane problem: paying for all of it. The company’s answer is tighter product focus. The key move is combining Codex, its coding agent, with ChatGPT in a project internally called The Merge. That effort produced ChatGPT Work, which is supposed to put agentic AI in front of a mass audience. Instead of just answering questions, it handles tasks. Thibault Sottiaux, the product lead running the project, told TIME that OpenAI is close to shipping something with persistence and always-on execution. It is definitely going to feel like a new thing to people.
OpenAI is also experimenting with new ways to make money. Ads inside ChatGPT have performed well enough in early tests that leadership is ramping up the volume. The ad revenue could help cover costs for the 92% of ChatGPT consumer users who do not pay for a subscription, according to CFO Sarah Friar. The company is also testing Sponsored Agents, a format where clicking an ad drops users into a brand-hosted AI experience.
What it means
For people making things, the shift is from asking questions to handing over tasks. Astra acts as a persistent agent that runs experiments and writes code without constant supervision. ChatGPT Work aims to handle actual workflows rather than just generating text. This changes the daily rhythm for users who rely on these tools, moving them toward a model where the software executes complex, multi-step actions on its own.
