OpenAI's Altman Says Astra Will Reach AGI-Level by Year's End

These remarks come from an in-depth TIME report by reporter Alex Heath, based on more than two weeks of interviews with over 20 OpenAI executives, employees, investors, and rivals.

Astra Can Do a Week's Worth of Research

At the center of this confidence is Astra, the model family OpenAI first teased last year as an automated AI research intern. Our earlier coverage, OpenAI's New Astra Model Flagged for Critical Cyber Capabilities, reported that the model had come close to a "critical" cyber-capability threshold in internal safety evaluations. This latest report reveals the other side of that same capability.

Chief Scientist Jakub Pachocki told TIME that Astra has already cleared internal benchmarks. Give it an experimental idea, and it will write the code itself within OpenAI's codebase, run the experiment, and report back the results. Hand it a single research paper, and it can complete what would take a human researcher a full week. Astra is also said to enable "persistent agents" — systems that can work on tasks autonomously over extended stretches of time.

Altman told customers in a demo, according to TIME, that this would be "the first time the model has meaningfully invented something new." That, he suggested, is what he means by an "AGI-like" moment.

The "80%" Figure — and a Definition Still in Dispute

Underlying these statements is a bet that AI may have already begun to recursively improve itself. Some researchers say that capability remains far off; others say they're already seeing early signs of it. Whether language models alone can make new discoveries and generalize from them is also still hotly debated — and that capability is widely treated as a baseline requirement for AGI.

The trouble is that "AGI" itself has no precise, agreed-upon definition. OpenAI's own charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work," a framing centered entirely on economic output. Critics counter that AGI requires a robust understanding of the world, and that large language models are just one piece of that puzzle. In other words, Altman's "AGI by year's end" claim only holds up if you accept his particular definition.

Devices, Agents, Ads: Solving the Money Problem Too

Alongside these lofty ambitions, OpenAI faces a much more immediate challenge: paying for all of it. The company's answer is to consolidate its products. At the center of that effort is an internal initiative called "The Merge," which combines the coding agent Codex with ChatGPT. The result is ChatGPT Work — a product built not just to answer questions but to actually carry out tasks. Thibault Sottiaux, who leads the project, told TIME the team is "close" to shipping something with "persistence and always-on execution."

According to Altman, OpenAI is also developing three categories of hardware: desktop, pocket, and wearable devices. TIME reports the first product will be a puck-shaped device that senses its surroundings and holds voice conversations, expected to launch in early 2027. Altman also said OpenAI will "definitely" build humanoid robots.

Revenue experiments are running in parallel. Early results from ads inside ChatGPT have been strong enough that the company is expanding the rollout — a meaningful move given that, according to CFO Sarah Friar, 92% of ChatGPT's consumer users don't pay for a subscription. OpenAI is also testing "sponsored agents," a format where tapping an ad takes users into an AI experience run by the brand itself.

Editor's Take

Altman's "AGI by year's end" claim shouldn't be taken at face value. Just months ago, this same company was internally warning that this same model, Astra, had come close to a "critical" cyber-capability threshold. That context makes today's confidence look like part marketing, part genuine concern — because research-intern-level automation and cyber-risk-level autonomy are two faces of the same underlying capability. It also suggests capability is advancing faster than safety controls can keep up.

Having spent the past few years putting GPT-series models to real-world use, I've noticed a pattern: every time a new model ships, the company insists "this one's different," but the practical gains tend to stay confined to narrow lanes like coding, summarization, and search. What OpenAI is emphasizing this time is an attempt to widen that lane into research automation itself. If Pachocki's claim holds — that a single paper can be turned into a week of human labor — this isn't just the next chatbot update. It's a challenge to the basic structure of R&D organizations.

For companies watching from outside the U.S., there are two things worth tracking. First, if an always-on agent like ChatGPT Work actually materializes, the current subscription model — where one chatbot gets shared across multiple teams — could shift quickly toward task-based billing. Second, the ad and sponsored-agent experiments are a signal that OpenAI is pushing ChatGPT's revenue model closer to that of a content platform, especially given that 92% of its users are on the free tier. Brand marketing teams should keep an eye on this shift starting now.

The things to watch in the coming weeks are whether Astra actually ships on schedule, and whether its safety evaluation results have eased. If a capability announcement arrives alongside a re-evaluation of safety, that would lend real weight to OpenAI's confidence. If the capability announcement comes first and the safety story lags behind, this "80%" figure will end up looking more like fundraising rhetoric than fact.