OpenAI's Astra Shifts AI Focus from Answering to Executing Long-Term Tasks

What’s truly remarkable about Astra isn’t how well it solves a single math problem, but that OpenAI is transforming cutting-edge models from “instant answer systems” into execution systems capable of operating for hours or longer, autonomously breaking down tasks and interacting with real software.

Author and source: 0x9999in1, ME News



• What’s truly remarkable about Astra isn’t how well it solves a single math problem, but that OpenAI is transforming cutting-edge models from “instant answer systems” into execution systems capable of operating for hours or longer, autonomously breaking down tasks and interacting with real software.
• Sixteen agents collaborating to prove a mathematical theorem, directly operating a computer, and the “AI Research Intern” demo all point to the same shift: the core metric of AI competition is moving from the quality of single responses to how long and how complex a workflow an AI can independently manage.
• If this capability can be stably deployed, the logic behind enterprises purchasing AI will change: in the past, they bought software licenses and Copilots; in the future, they may increasingly move toward purchasing “digital labor” and task throughput.
• But a closed demonstration is not a production environment. Astra has not yet been publicly released, and the so-called “research intern” is an internal OpenAI standard; the mathematical demonstration yields candidate proofs, not independent verification of their scientific validity.
• More concerning, capability improvements have begun to constrain OpenAI’s own development speed. In August, OpenAI acknowledged that Astra may have reached its highest-level “Critical” cybersecurity capability threshold, leading to a pause in certain training and evaluation activities.
• So, whether Astra is AGI is now a debate with limited significance. The truly important question has become: What will happen to the economic models of businesses, labor markets, and AI companies when AI first gets the chance to move from “answering your questions” to “continuously doing work for you, even helping to research the next generation of AI”?

On the surface, Astra's private preview could easily be interpreted as just another round of hype before a model launch.

Dozens of executives from OpenAI’s key clients arrived at the newly opened MB0 building in San Francisco. Outside, protesters with “Stop the AI Race” signs still gathered; inside, a full wall of greenery was positioned against the glass doors, blocking the view between the two sides. Sam Altman had just returned from Washington, where he had previously briefed U.S. government officials on the yet-to-be-released family of models.

This scene actually captures the state of today’s AI industry more accurately than any keynote: capital, customers, and governments are all waiting for capabilities to keep advancing, while another group is already worried it’s moving too fast.

But what truly matters about Astra is not "how many more percentage points the model has improved."

The first demonstration on stage involved 16 AI agents tackling a research-level mathematics problem. They autonomously broke the problem into multiple subproblems, attempted solutions individually, coordinated their results, and combined them into a candidate proof. It is essential to clarify the term “candidate”: this is not a newly verified mathematical theorem, nor does the simultaneous operation of 16 agents automatically equate to a research team of 16 accomplished mathematicians. Multiple models could easily share the same error, and increasing the number of agents does not inherently guarantee a linear improvement in accuracy.

But it is still important.

In the past, when we evaluated large models, we were accustomed to focusing on one question: Could it answer it?

Astra is trying to reframe the question: Can it organize its own work and carry out a complex task continuously?

The two may seem to differ by only one step, but their business implications could differ by an entire era.

The earliest typical interaction with ChatGPT was for humans to ask questions and for the machine to respond; later, Copilot began integrating into programming, office work, and creative workflows, where humans continuously give instructions and AI provides ongoing assistance. Agents take this a step further, meaning humans only need to state their goal, while the machine determines the steps, selects tools, checks intermediate results, and proceeds autonomously.

So the second demonstration is even more worth serious attention from corporate executives than a math problem.

Astra directly operates the computer, switching between multiple common desktop applications, creating and modifying content, and completing tasks at an extremely fast pace. Altman described watching the model use the computer in a “superhuman, very fast” way as one of the most astonishing things OpenAI employees have experienced recently.

Why is operating a computer so important?

Because most work in the corporate world isn’t a neat question you can plug into a chat box.

Real work is scattered across browsers, Excel, email, code repositories, ERP, CRM, design software, internal databases, and approval systems. An employee’s value has never been just about “knowing the answer,” but about moving information between a dozen systems, making judgments, modifying files, driving processes, and ultimately getting things done.

As long as AI cannot cross this threshold, it remains primarily a knowledge tool.

Only after crossing over does it truly begin to approach "labor."

Altman described the core product vision for Astra as persistent agents—agents that exist continuously and work over the long term.

This term may not sound as exciting as AGI, but it could be closer to what will truly impact corporate profit statements in the coming years.

OpenAI's chief scientist Jakub Pachocki told TIME that Astra has met the company's internal standard for an "automated AI research intern": given an experimental idea, it can enter OpenAI’s own codebase to implement the solution, run the experiment, and return results; given a research paper, it can perform follow-up work that previously might have taken a human researcher about a week.

This is still only an internal evaluation by OpenAI, without independent third-party validation, and cannot be simply interpreted as "Astra can already replace researchers." However, the signal it sends is clear enough: OpenAI's unit of measurement is shifting from tokens, responses, and conversations to task duration.

This direction was not suddenly created by OpenAI itself.

The independent evaluation agency METR has spent the past few years attempting to measure Agent capabilities using a “task-completion time horizon”: instead of asking whether a model can solve a specific type of problem, it observes how long it takes the model to complete tasks with a certain success rate—equivalent to the time a human expert would need.

Even after the 2026 update, METR's data still shows a clearly exponential improvement trend. In the long term, the software and research task durations for frontier models approximately double every 6 to 7 months; since 2023, the pace has been even faster. By May of this year, some of the latest publicly available models have already approached or exceeded the effective measurement range of around 16 hours in their existing benchmarks, prompting METR to specifically caution that data exceeding this length should not be overinterpreted.

This precisely explains why Astra emerged.

As the tasks the model can reliably complete extend from minutes to hours, and then from hours to a full day or longer, the product form will inevitably change. You no longer need to stay by the chat window to correct it; instead, you begin to assign tasks like you would to a colleague: “Figure out this issue, test three approaches, organize the data, and give me the results tomorrow.”

The so-called "virtual colleague" has never truly been defined by whether it can speak like a human.

But whether it will keep working after you leave.

This is also the financial implication of Astra that is most easily underestimated.

Over the past three years, the dominant model for enterprises purchasing generative AI has essentially been an extension of the traditional SaaS model: adding one AI seat per employee, with an additional cost of dozens of dollars per month. This applies to Microsoft Copilot, ChatGPT Enterprise, and numerous other AI SaaS products.

But as the persistent agent matures, this pricing logic will become increasingly awkward.

Why?

Enterprises are no longer truly purchasing "permission for an employee to use AI," but rather the ability of AI itself to accomplish work.

Suppose a finance team previously required 10 people to handle data organization, draft reporting, anomaly detection, and cross-system data entry. In the future, the question managers face may no longer be “Should I buy a Copilot for each of these 10 people?” but rather “How much Agent workload do I need to purchase to complete these tasks?”

At this point, the AI's competitor is no longer just software budgets.

It is beginning to enter the personnel budget.

Data from the 2026 Stanford AI Index has already revealed this contradiction. On one hand, enterprise AI adoption continues to rise rapidly: in 2025, 88% of organizations surveyed were using AI in at least one area, and 70% were using generative AI; on the other hand, the proportion of enterprises that have actually deployed agents remains in single digits across the vast majority of business functions.

In other words, people are already very willing to "use AI," but have not yet widely become willing to "hand things over to AI."

This is precisely the gap Astra is trying to bridge.

Once crossed, the impact will not be confined within tech companies. Stanford data shows that since 2024, employment among software developers aged 22 to 25 in the U.S. has declined by nearly 20%; meanwhile, about one-third of surveyed companies expect AI to lead to workforce reductions over the next year, yet the overall job market has not yet seen a comparable scale of widespread unemployment.

This suggests that the impact of AI on the labor market is unlikely to be like flipping a single master switch, but rather will first alter the marginal demand for hiring channels, entry-level positions, and standardized knowledge work.

Precisely for this reason, the term "AI Research Intern" is actually quite nuanced.

Today, it might first handle the time-consuming but clearly defined experimental tasks for researchers; tomorrow, companies will ask: Where’s the financial analyst’s initial modeling? Where’s the consultant’s information gathering? Where’s the programmer’s routine development? Where’s the market team’s cross-system execution?

In the past, we have been discussing whether AI will "replace a profession."

A more realistic process is that AI will first gradually take over the most standardized, easiest to verify, and time-consuming modules within a profession. The job won’t disappear overnight, but the work that once required ten new hires at a company may gradually need only six, then four people, plus a team of agents.

What truly transforms the job market is not usually "AI suddenly being able to do everything."

Instead, companies have found: the next employee doesn’t need to be hired yet.

If the story ended here, Astra would simply be a product with tremendous commercial imagination.

The problem is that OpenAI itself has already shown that things are not this simple.

On August 7, OpenAI publicly stated that, according to its latest internal assessment, the company can no longer rule out Astra reaching the "Critical" cybersecurity capability level in its Preparedness Framework.

This threshold is not "the model can write code."

According to OpenAI's own definition, a Critical level means the model could, without human intervention, identify and develop zero-day vulnerabilities of varying severity in large-scale, hardened real-world critical systems, or autonomously design and execute end-to-end novel attack strategies against hardened targets based solely on a high-level objective.

OpenAI emphasized that this level is currently only not ruled out, rather than confirmed that Astra possesses all the aforementioned capabilities; additionally, Astra is unrelated to OpenAI’s previous security incident involving Hugging Face.

But the company's response has been enough to speak for itself.

OpenAI has strengthened network isolation, sandboxing, model weight protection, and end-to-end monitoring, suspending Astra activities that do not meet the new security standards. On August 18, the company further disclosed that the latest reinforcement learning training for deploying new models was paused for two weeks to harden the environment and expand monitoring coverage, with the largest-scale frontier RL training still suspended, and much of the training and evaluation work involving Astra yet to be resumed.

Even security itself is beginning to become an expensive computational expense.

OpenAI estimates that this monitoring system may currently consume an additional 20% of computational power relative to the monitored inference workload.

This set of numbers is well worth remembering in the capital markets.

The story in the AI industry over the past few years has been: the stronger the model, the greater its commercial value, making it worthwhile to invest more in GPUs, data centers, and electricity. But Astra is now revealing another cost curve—stronger models don’t just become more expensive to train; security, isolation, monitoring, and governance themselves begin to consume significant compute resources.

This means the economic model of the next generation of AI cannot be based solely on "cost per million tokens."

Uncontrolled risk, security redundancy, manual approval, permission isolation, and real losses caused by errors must also be calculated.

A model that helps you write emails going wrong might only cause embarrassment.

An agent with long-term memory, code execution capabilities, and network access, capable of autonomously operating enterprise systems for hours, makes an error of a completely different nature.

The closer the capability is to that of an "employee," the more the access control must align with "employee" standards; once capability exceeds that of a regular employee, security requirements must even exceed those for employees.

So what Astra truly challenges OpenAI on may not be “whether it can be done.”

But can such a system be turned into a product that companies dare to deploy in production environments?

The furthest Altman went in the private preview was his prediction that Astra might become the first model capable of "inventing something new in a truly meaningful way," which he called "something very much like AGI."

This statement certainly has a strong Altman style.

There is currently no public evidence demonstrating that Astra has consistently generated significant, peer-validated scientific breakthroughs, nor has OpenAI released sufficient external test results. Therefore, labeling Astra as “AGI has arrived” is both unnecessary and unsupported by facts.

But why Altman specifically emphasized "discovering new knowledge" is worth serious study.

Because this is where the AI industry will truly experience exponential change.

If AI merely assists humans in improving efficiency in coding, design, and document analysis, then even the most aggressive advancements in the industry are essentially a revolution in productivity tools. It can greatly enhance economic efficiency, but still largely relies on humans to determine what to research next, how to design experiments, and how to create the next generation of technologies.

If AI can participate in AI research itself, the logic changes.

An agent capable of reading papers, proposing experimental designs, modifying large codebases, running experiments independently, and analyzing results—even if it only reaches the level of an excellent research intern—can already significantly expand the effective research capacity of a top-tier lab.

Today, 100 researchers may be able to run 500 agents behind the scenes; in the future, the number of experiments each researcher can advance simultaneously each day may no longer be limited by the 24 hours in a day, but rather by computational power, experiment cost, and verification capacity.

Thus, a concept that was once very abstract—“AI helping to develop better AI”—began to take shape.

This is the aspect of Astra that deserves the most attention—and also requires the most restraint.

There is still a vast gap between “completing experiments designed by others” and “proactively proposing valuable new research directions”; the gap is even greater between “generating a seemingly reasonable new idea” and “producing reproducible, verifiable, and truly scientifically advancing new knowledge.”

Sixteen agents working simultaneously does not mean sixteen times the innovation.

Running a hundred thousand experiments per day doesn't mean you'll naturally achieve a breakthrough.

Science never lacks hypotheses; what's truly scarce is knowing which questions are worth asking and which results truly matter.

No one can draw a conclusion from a closed-door demonstration as to whether Astra has crossed this line.

So, discussing whether Astra is an AGI is somewhat premature.

The definition of AGI is inherently highly ambiguous. OpenAI has long defined it as a highly autonomous system that outperforms humans in most economically valuable work; other labs, scholars, and the public have entirely different criteria for judging it.

Instead of arguing over a label, observe three more specific issues.

First, how long can Astra independently handle tasks, and what is its reliability rate?

Second, how much time do humans need to spend reviewing and correcting errors once it enters enterprise production? If an Agent works for eight hours but requires engineers to spend another three hours verifying its output, its true productivity gain is entirely different from what’s shown in demo videos.

Third, and most importantly: can it advance automated research from “executing experiments” to “discovering genuinely valuable new knowledge”?

The first two questions determine whether Astra is a huge business.

The third question determines whether it is a turning point of an era.

From this perspective, the wall of greenery pushed up at the MB0 entrance carries symbolic weight. On one side are dozens of client executives witnessing, for the first time, digital agents that may work for them long-term; on the other, protesters demanding the entire industry halt its AI race. Both sides hold radically different visions of the future, yet both acknowledge the same thing: this technology is becoming more powerful than ever before.

Astra has not been publicly released, and its internal capabilities have not been sufficiently validated externally. It may be delayed, expose numerous reliability issues in real-world environments, or ultimately fall far short of Altman’s description of AGI.

All of these possibilities should be retained.

But one thing has become increasingly clear.

In the previous stage of large model competition, the focus was on who could answer questions the world already knew; the next stage, represented by Astra, begins to ask who can consistently take action, independently complete complex tasks, and ultimately help humanity explore questions the world has yet to answer.

Between chatbots and virtual colleagues lie four high walls: reliability, security, cost, and organizational trust.

What OpenAI is truly betting on now is that Astra can overcome each challenge one by one.

But if it truly flips, when people look back at 2026, what really matters may not be how many points one model led on a leaderboard.

Instead, machines are beginning for the first time to acquire a capability that was previously primarily human: receiving a task and then carrying it out independently.

1. Alex Heath, “Inside OpenAI’s Reboot,” TIME, August 26, 2026 (republished by Yahoo Tech).
2. OpenAI, “Responding to the next frontier of critical cyber capabilities”, August 7, 2026. (OpenAI)
3. OpenAI, “Pacing model development in an era of cyber-critical capabilities”, August 18, 2026. (OpenAI)
4. Cat Zakrzewski, Ian Duncan, Riley Beggin, “Sam Altman courts Washington as OpenAI pushes a powerful new AI”, The Washington Post, July 31, 2026. (The Washington Post)
5. Stanford Institute for Human-Centered AI, The 2026 AI Index Report — Economy, 2026. (斯坦福HAI)
6. METR, “Task-Completion Time Horizons of Frontier AI Models”, updated on May 8, 2026. (METR)
7. METR, "Time Horizon 1.1", January 29, 2026. (METR)
8. Stop The AI Race, “Occupy OpenAI / Stop The AI Race”, 2026. (stoptherace.ai)