OpenAI's Secretive Astra AI Can Work for Days and Invent New Knowledge as Sam Altman Predicts AGI by 2026

OpenAI's experimental Astra system can work for days at a time, operate computers at what insiders describe as 'super-human' speed and has already hit the company's internal benchmark for an automated research intern.

At the same time, chief executive Sam Altman says he expects OpenAI to have an internal system he would call artificial general intelligence (AGI) by the end of 2026.

Astra Can Work for Days

OpenAI demonstrated Astra to customers and executives at its headquarters this month, including a test in which 16 AI agents worked together on a research-level mathematics problem.

The agents divided the problem into smaller tasks before combining their work. Astra can also navigate desktop software and operate across applications, with Altman describing the technology as capable of running 'persistent agents' that continue working rather than stopping after a single response.

Altman said he expects Astra to become the first model that 'invents new things in a way that matters'. He described that capability as 'a very AGI-like thing'.

OpenAI Says It Is 80% of the Way to AGI

OpenAI chief research officer Mark Chen said the company was '80% of the way' to AGI. Co‑founder Greg Brockman said the period could eventually be remembered as the point when AGI was created.

Altman put a clearer timeframe on the prediction, saying he expected the company to have an internal system he would call AGI by the end of 2026. That is an assessment of an internal system, rather than a formal announcement that OpenAI has achieved AGI.

It also does not establish that OpenAI will release a public AGI system in 2026. The company has not said that Astra already meets its full definition of AGI.

Astra Has Passed an Internal Research Benchmark

OpenAI chief scientist Jakub Pachocki said Astra already meets the company's internal benchmark for an automated research intern.

The system can take an experimental idea, implement it in OpenAI's codebase, run the experiment and return the results.

It can also take a research paper and carry out follow-up work that would previously have taken a human researcher about a week.

Those capabilities have been assessed internally and have not been independently verified.

Cybersecurity Concerns Have Slowed Development

Astra's capabilities have raised a separate concern inside OpenAI. On 7 August, the company said it could not rule out the model reaching its 'Critical' cybersecurity capability threshold.

OpenAI subsequently introduced tighter security controls, isolated testing environments, restricted network and tool access and expanded monitoring. On 18 August, it said some Astra training and evaluation workloads remained paused while those safeguards were strengthened.

The company's Preparedness Framework uses the Critical threshold for capabilities that could enable serious cybersecurity attacks, including the development of functional exploits against hardened systems.

OpenAI's 2028 Goal Has Shifted

The year‑end AGI prediction comes only months after OpenAI published a longer timeline for autonomous research. In June, the company said it expected AI systems to conduct a significant fraction of its internal research by March 2028.

That plan referred to an automated AI researcher rather than an internal system Altman would label AGI. The change reflects the rapid development of systems designed to perform increasingly complex research with less human involvement.

OpenAI's AGI Definition Is Broad

OpenAI's Charter defines AGI as highly autonomous systems that outperform humans at most economically valuable work. A model that can solve advanced mathematics problems, operate a computer or conduct research does not automatically meet that standard.

Altman's prediction depends on how OpenAI evaluates its own internal system against that definition. It is not an independent determination that AGI will arrive in 2026.

Astra also remains unreleased, with its capabilities described through internal testing and demonstrations. The company must still determine whether its performance and safeguards are sufficient for wider deployment.