Loading page…
The target depends on an economics-based definition of general intelligence and company-described research performance, while the field still lacks a shared technical finish line.
Listen to this story
OpenAI’s AGI roadmap is being tested against a nearer product question: whether Astra can operate as a persistent research agent. Chief research officer Mark Chen puts the company 80% of the way to its internally defined threshold—systems outperforming humans at most economically valuable work—while Altman has not said it is achieved. Astra is expected to code, run experiments, and handle week-long research tasks; ChatGPT Work extends task execution to consumers.
Astra is designed to turn a research idea into code, execute experiments in OpenAI’s codebase, and return results.
Pachocki says Astra can tackle paper-based tasks that previously took a human researcher about a week.
OpenAI’s consumer agent push includes The Merge, which combines Codex with ChatGPT to produce ChatGPT Work.
Sam Altman says OpenAI could have an internal system he would call artificial general intelligence by the end of 2026. The claim turns on OpenAI’s own definition of AGI and Astra, an upcoming model family the company describes as capable of extended research work; neither resolves the underlying dispute over what the label should require.
OpenAI defines AGI as highly autonomous systems that outperform humans at most economically valuable work. The term has no precise technical definition, and researchers disagree over whether language-model-based systems can make discoveries and generalize in the ways AGI would require.
Altman has said OpenAI has not reached that point yet. Mark Chen, the company’s chief research officer, estimated that OpenAI is 80% of the way to AGI—an internal measure, rather than an independently agreed finish line.
I expect this will be the first model where the model actually invents new things in a way that matters.
Sam Altman, on Astra’s potential impact, as quoted by TIME
Astra is intended to function as an automated AI research intern. Chief Scientist Jakub Pachocki said it can take an experimental idea, write code in OpenAI’s codebase, run the experiment and return results. He also said it can work from a research paper on tasks that would have taken a human researcher about a week.
Those capabilities describe an agent that can execute a research loop, not necessarily one that produces important discoveries. Astra is also intended to support persistent agents—virtual coworkers that work autonomously on tasks over longer periods. Whether that work reliably generates new knowledge remains the material distance between OpenAI’s described system and Altman’s AGI framing.
OpenAI is combining Codex with ChatGPT in an internal effort called The Merge, which has produced ChatGPT Work. The stated product shift is from answering questions toward handling tasks, putting the company’s agent ambitions before a broader audience.
The nearer test is narrower than the AGI label: whether OpenAI can ship persistent agents that complete valuable work over sustained periods. A working research agent would strengthen its case, but would not settle the argument over whether that performance is general intelligence.
Loading discussion...
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.