Sam Altman puts a timeline on AGI

Sam Altman just did what tech executives usually avoid at all costs: he put an actual deadline on AGI.

In a sweeping profile published by TIME, Altman claimed that an artificial general intelligence matching his personal definition will be running inside OpenAI before the year is out. That means an internal system capable of meeting or beating humans across virtually every task is, in his eyes, just months away.

An automated researcher named Astra

Altman wasn't the only one painting an aggressive picture. Other OpenAI leaders backed him up with unusually specific progress markers.

Research lead Mark Chen estimated that the team is already about 80% of the way to AGI. Co-founder Greg Brockman added that history will likely look back at this exact period as the moment AGI actually arrived.

Chief scientist Jakub Pachocki provided the most concrete proof point. He pointed to an internal model called Astra, which can take an academic paper and conduct a full week of solo researcher work on its own. That milestone satisfies OpenAI's 2026 automated-intern goal.

Altman described Astra as the first model that actually invents new things in a meaningful way, arguing that the creation of novel knowledge is inherently AGI-like behavior.

Yet the interview wasn't all triumphalism. TIME also dug into a security breach OpenAI suffered in July, which the lab followed up with a detailed internal report. As a result, OpenAI is attempting to refocus on safety and tap the brakes on frontier development.

Why it matters

We think the real story here isn't the flashy AGI label, which everyone defines differently anyway. It's the prospect of Pachocki's automated researcher.

If Astra can genuinely ingest research papers and churn out a week's worth of solid experimental work solo, the economics of R&D completely change. For builders and scientists, an automated intern that generates real, novel knowledge would collapse the time it takes to explore hypotheses from months to days.

That said, nobody outside OpenAI should expect to play with this anytime soon. Altman was careful to specify that the model will exist internally by year-end, which leaves a massive gap before any public release, especially given the cost and compute required.

There is also an awkward contradiction here. OpenAI just suffered a serious breach in July and publicly pledged to slow down frontier work to shore up security. Hyping an internal AGI milestone while nursing fresh security scars looks like an aggressive race for narrative dominance, and it leaves an open question about whether their internal guardrails can keep up with what Astra can invent.