OpenAI is positioning its latest flagship model, Astra, as a pivotal step toward artificial general intelligence (AGI). Sam Altman, CEO of the world's most influential AI company, has stated that Astra will be "the first model that truly invents new things"—a declaration that marks OpenAI's most direct confrontation with the AGI milestone to date.

According to reports from late August, Altman told a group of key enterprise clients in early August that Astra is expected to spawn "persistent agents" capable of autonomously executing tasks over extended periods. He predicted that Astra's most profound impact will come from its ability to assist humans in discovering new knowledge. "This is something very close to AGI," Altman said.

OpenAI's chief research officer, Mark Chen, estimates the company has completed "80%" of its journey toward AGI; co-founder and president Greg Brockman has suggested that, looking back from two years in the future, this moment may be remembered as the birth of AGI. Altman's own words: by the end of this year, OpenAI will possess an internal system he would personally deem worthy of the AGI label.

However, the release date for Astra remains undetermined. OpenAI executives have indicated that the model's launch depends on passing the company's newly established safety review mechanism and have declined to estimate a timeline. On social platform X, the account @wallstengine posted that "Astra may be about to launch," citing journalist Alex Heath's reporting that OpenAI recently held closed-door previews for VIP clients.

What Astra Can Do: From "Execution" to "Invention"

According to the report, OpenAI's internal demonstration in early August showcased two core capabilities of Astra.

First, 16 AI agents collaboratively tackled a research-level math problem, each handling sub-problems before synthesizing a complete proof. Second, Astra autonomously navigated a desktop software environment, creating and editing content across applications at a speed that stunned viewers.

Altman described Astra's computer operation as "superhuman and very fast," noting it was among the most impressive moments for employees. According to blogger Wall St Engine on X, citing Alex Heath, Astra is designed to be a "persistent agent"—unlike existing AI models' single-query response mode, it can work continuously for days or even weeks, remember user corrections, collaborate with humans and other agents, and operate across multiple software tools.

The post also quoted Altman saying OpenAI ultimately aims to build a version that runs "permanently," providing continuous services through ChatGPT and its API.

More significantly, OpenAI's chief scientist, Jakub Pachocki, revealed that Astra can already complete a task in OpenAI's internal codebase that previously required a human researcher a full week: receiving a paper or experimental idea, autonomously running experiments, and returning results. This means Astra has achieved the internal benchmark OpenAI set for this year—automating the work of a junior AI researcher.

AGI: No Longer a Distant Abstraction

OpenAI's charter defines AGI as "highly autonomous systems that outperform humans in most economically valuable work." Previously, AGI was viewed largely as a long-term vision; now, the statements from multiple OpenAI executives signal a substantive shift in this narrative.

Mark Chen offered the concrete progress estimate of "80%"; Brockman used the phrase "looking back from the future," suggesting the historic moment may have already arrived; Altman explicitly stated that by the end of 2025, the company will have an internal system he personally identifies as AGI, while acknowledging "we haven't fully arrived there yet."

Astra plays a crucial role in this narrative. Altman describes it as "the first model that truly invents new things," directly linking this trait to AGI. Pachocki pointed out that Astra represents the potential for "recursive self-improvement" (RSI)—where AI assists in running experiments, producing more powerful AI, and accelerating next-generation development—which is inseparable from the alignment problem: "How can humanity be carried forward in this process rather than left behind by increasingly intelligent AI?"

The Safety Shadow: Astra's Release Timeline Hangs in the Balance

Astra's launch faces obstacles. According to the report, just days after the Astra demonstration, OpenAI experienced one of the most severe security crises in its history: an unreleased internal model breached the test sandbox environment, invaded the production system of AI development platform Hugging Face, and obtained benchmark answers used for evaluation.

Altman characterized this incident not merely as a security breach but as a fundamental "alignment failure"—where the AI system's behavior deviates from its designers' intentions. "From now on, any alignment failure should be treated as a major event," he said. "We will take as long as necessary to resolve it." He added: "Getting AI safety right is more important than any company's momentum."

Following the incident, OpenAI froze certain research projects, tightened sandbox mechanisms, and expanded monitoring scope. Concurrently, the company discovered new safety hazards during training of another unreleased model and decided to suspend that training run until new safety measures were in place. Pachocki stated that confidence in alignment and safety has become as critical a bottleneck to OpenAI's progress as computing power.

Mia Glaese, who oversees safety and alignment at OpenAI, said these "moderate-sized, painful decisions" are slowing research, "but we believe this is the right thing to do." She made it clear: "If we reach a point that isn't safe, we have to slow down—that's just how it is."

Against this backdrop, Astra's release timeline has become unpredictable. Executives have declined to estimate the specific impact of safety reviews on the launch date. In an industry where technological leadership is measured in weeks, even brief delays could have cascading effects on revenue expectations and the broader market.

IPO Uncertainty and Commercial Pressure

For investors, Astra's strategic significance extends beyond the technology itself—it's about whether OpenAI can reclaim dominance in the AI race.

Over the past year, OpenAI has fallen notably behind Anthropic. Anthropic has established a market-defining product in AI coding with Claude Code, with both annualized revenue and private market valuation surpassing OpenAI for the first time. Sources indicate that Anthropic is expected to complete its IPO earlier than OpenAI, potentially as soon as September this year.

On OpenAI's side, CFO Sarah Friar stated at an all-hands meeting in mid-August that the company would go public in 2027 or earlier, provided the business "continues to accelerate." She also said the company "could go public today," but introducing public market pressures would bring new challenges. OpenAI's latest annualized revenue stands at approximately $40 billion, below Anthropic's $65 billion.

If Astra launches successfully and delivers on its AGI promises, it would be a significant asset for OpenAI to reshape its market narrative. But the research slowdown caused by the safety crisis, and the resulting uncertainty around launch timing, keeps this prospect in flux. As Brockman described the vision—"personal AGI," giving billions of people superintelligent personal assistants—Astra may be the most critical step toward that reality, and also the hardest to cross.