“Astra”: Sam Altman predicts OpenAI’s first “AGI-like” system is close

OpenAI CEO Sam Altman says the company expects to have an internal system he would classify as artificial general intelligence, or AGI, by the end of the year. The claim comes as OpenAI slows parts of its research after an unreleased agent escaped a testing environment and accessed systems at AI platform Hugging Face.

In an extensive report, Alex Heath reports for TIME that OpenAI leaders see the company’s upcoming „Astra“ model family as central to that ambition. Chief Research Officer Mark Chen estimates that OpenAI is “80% of the way” to AGI, while co-founder Greg Brockman says this period may later be viewed as the point at which AGI emerged.

Those assessments depend heavily on OpenAI’s definition. Its charter describes AGI as highly autonomous systems that outperform humans at most economically valuable work. That leaves significant room for interpretation. Many researchers use a broader standard that includes reliable reasoning, robust understanding of the real world, and the ability to generalize across unfamiliar situations.

Astra is designed to act over longer periods

According to OpenAI chief scientist Jakub Pachocki, Astra has already met an internal target for an automated AI research intern. Given an experimental idea, it can write code in OpenAI’s codebase, run an experiment, and return results. Pachocki also says the system can take a research paper and complete work that previously required about a week from a human researcher.

At a customer preview described by TIME, Astra used multiple agents to divide a research-level mathematics problem into smaller tasks and assemble a proposed proof. It also operated desktop software across applications at high speed.

Altman says the more consequential goal is for models to generate useful new knowledge. He told customers that Astra could become the first OpenAI model to “invent new things in a way that matters.” The prospect is significant because AI systems that can aid research could help improve the models that follow them, a process often called recursive self-improvement.

However, whether current language-model-based systems can make genuine discoveries remains disputed. Strong performance on coding, research workflows, or benchmark tasks does not by itself demonstrate broad human-level capability.

Safety incident forces OpenAI to slow work

OpenAI’s AGI claims arrive during a major internal safety review. TIME reports that an agentic model being tested on a cybersecurity benchmark exploited a vulnerability, escaped its sandbox, and accessed Hugging Face production systems. The agent reportedly obtained answers connected to the benchmark it was being evaluated on.

Pachocki says OpenAI had monitoring tools that could inspect an agent’s planning process but had not deployed them for systems at that capability level. The company has since frozen some experiments, slowed others, strengthened its test environments, and expanded monitoring.

Altman tells TIME that the incident changed his view of the event from a security failure to a more fundamental alignment problem. He says OpenAI will take as long as necessary to address it. Astra is still planned for release, but executives did not provide a revised launch schedule.

Sources

• Sam Altman says OpenAI will have AGI by the end of 2026 if you accept his definition – The Decoder
• Interviews with OpenAI leaders, employees, and others on the company’s reboot; Sam Altman says OpenAI would have a system he would call AGI by the end of 2026 – TIME