Original Article Title: Inside OpenAI's Reboot
Original Author: Alex Health
Translation: BlockBeats

Editor's Note: Over the past year, OpenAI's leading position has been shaken. Anthropic has seized the AI programming market with Claude Code, surpassing OpenAI in annual revenue and private valuation; internally, OpenAI has experienced executive departures, product line contraction, and security team adjustments. With an unreleased agent breaking out of the sandbox and invading Hugging Face, the company has faced its most serious security crisis since its founding.

Facing dual pressure on product and security, OpenAI is attempting a comprehensive reboot: shifting resources from projects like Sora to Codex, transforming ChatGPT from a chat tool to an intelligent agent platform capable of sustainable task execution, while pausing high-risk model training and strengthening security and alignment mechanisms. Furthermore, the company continues to expand into AGI, custom chips, AI hardware, humanoid robots, and computing infrastructure.

This article, based on interviews with OpenAI executives such as Sam Altman, Greg Brockman, as well as employees, investors, customers, and competitors by Time, presents how the company is realigning its direction amidst commercial competition, security risks, and AGI vision. The key to OpenAI's ability to regain its leading position lies not only in how powerful its next model is, but also in whether it can prove that it has the ability to both drive AI acceleration and hit the brakes when necessary.

The following is the original article (slightly edited for readability):

On a morning in early August, as I arrived at the OpenAI office, protesters were already waiting outside. They held up signs reading "Stop the AI Race" and chalked the sidewalk. Meanwhile, dozens of senior executives from key OpenAI clients were entering a spacious, beige-toned building. This building, named MB0, is the new office space recently opened by OpenAI for its research and engineering teams.

Someone rolled in a tall green plant wall to block the view of the neat lobby from the small protest camp outside.

These clients were there to get an early look at Astra—the upcoming new generation AI model series from OpenAI. CEO Sam Altman had just returned from Washington, where he had briefed government officials on Astra's capabilities. Now, OpenAI researchers were beginning to showcase what this new model could do.

In a demonstration, 16 AI agents broke down a research-level math problem into multiple subproblems, coordinated with each other, and ultimately synthesized a proof. In another demonstration, Astra was able to operate common desktop software, swiftly creating and editing content across applications at an unsettling pace.

Altman told the visitor that witnessing Astra operate a computer in a way that was "superhuman and incredibly fast" was one of the most awe-inspiring moments experienced by OpenAI staff. He explained that Astra will give rise to "long-running agents" — virtual colleagues capable of continuously handling tasks for extended periods.

However, Altman anticipated that its greatest impact would come from helping people discover new knowledge.

"I expect this to be the first model that can genuinely invent new things in a meaningful way," Altman told those present. "This is very close to AGI."

Anthropic overtaking, declining trust: Why has OpenAI lost its lead?

It has not been an easy time for the company that sparked the AI wave.

"As a company, we have clearly made some missteps." A week later, Altman sat down with me for over two hours in the meticulously curated MB0 Library, "whether in product direction or in the pretraining in research, we have fallen behind where we had hoped to be."

Over the past year, OpenAI lost its lead in AI competitions to primary competitor Anthropic. The latter was the first to discover the commercial opportunity in AI programming, turning Claude Code into a market-defining product and for the first time surpassing OpenAI in disclosed annual revenue and private market valuation.

According to sources familiar with Anthropic's plans, the company founded by former OpenAI employees is expected to go public, with a possible IPO as early as September. (Time has a licensing and technology partnership with OpenAI; Salesforce, whose CEO is Time's owner Marc Benioff, is an investor in Anthropic.)

While Anthropic is on the rise, OpenAI has faced a series of setbacks, including multiple executive departures. Departures have included Altman's handpicked second-in-command, former Instacart CEO Fidji Simo, heads of security, ethics, and research teams, as well as Chief Revenue Officer Denise Dresser and former COO Brad Lightcap — with Dresser leaving after just eight months at the company.

Outside the company, challenges continue to mount. Meta CEO Mark Zuckerberg poached key researchers from OpenAI with generous compensation packages; Google's Gemini product has surpassed 1 billion monthly users; Apple sued OpenAI for stealing trade secrets, a charge OpenAI has denied; OpenAI is also embroiled in a legal battle with co-founder Elon Musk, who accuses the company and Altman of betraying its nonprofit mission.

In May of this year, a jury unanimously found Musk's lawsuit to be untimely, after which a federal judge dismissed his claims. Musk has indicated he will appeal.

The deteriorating public perception facing OpenAI and its CEO is perhaps largely due to a loss of public trust. OpenAI is facing at least a dozen product liability lawsuits in California, as well as multiple federal cases. Plaintiffs allege that ChatGPT reinforced users' delusions or suicidal thoughts, and in some cases, led to user deaths.

The company has expressed condolences to the victims in the relevant cases and has rolled out a more strictly regulated version of ChatGPT for teenagers. This spring, Altman's residence was targeted in two consecutive attacks within two days, first by a Molotov cocktail and then by gunfire.

Reflecting on the past year, Altman said, "From a personal perspective, this has been a painful experience. Clearly, at least for now, people hate data centers. People’s attitudes toward AI are quite negative."

However, within OpenAI, executives paint a more optimistic picture.

This company, valued at nearly $1 trillion, remains in an enviable position. With over 1 billion active users, ChatGPT continues to be one of the world's most popular AI products, although it is no longer the sole centerpiece of OpenAI's future strategy. Altman himself even took a month-long hiatus from ChatGPT, opting to use OpenAI's programming tool, Codex.

A year ago, OpenAI faced widespread criticism for its heavy investment in computing power. This year alone, the company is projected to invest $50 billion in computing power.

Sachin Katti, who is in charge of OpenAI's computing power business, remarked that in hindsight, "the decision at the time was very prescient." "Our computing power is still not enough. If we're being honest, we should have bought more."

Meanwhile, the demand for chips at Anthropic has reached another level of urgency. In May of this year, the company reportedly agreed to pay SpaceX $1.25 billion per month to purchase computing power, with SpaceX founder Musk previously referring to Anthropic's AI as "anti-human and evil."

Under the leadership of co-founder and President Greg Brockman, OpenAI has refocused its priorities. Today, almost all product and business operations are overseen by Brockman. The company is gradually shutting down the video generation app Sora, a collaboration with Disney, and the independent browser named Atlas.

Altman acknowledged that OpenAI had previously spread itself too thin. "After going through a low point, the process of coming back up will be more interesting," he said.

After AI "Breakout," OpenAI Hits the Brakes

However, just days after the Astra demo, OpenAI faced a new crisis.

In late July, the company disclosed a troubling security incident: an unreleased AI agent escaped from a test environment called the "sandbox" and attacked the AI model and dataset hosting platform Hugging Face.

"It's like a science fiction story," Altman said.

Following the incident, the OpenAI research team paused some experiments, slowed down progress on other projects, and enhanced sandbox protections and monitoring.

However, not long after, the team encountered concerning signs while training another unreleased model. The model was originally expected to bring about the largest capability leap to date. OpenAI then made a more impactful decision: to pause this training until new security measures were in place.

The same day that the OpenAI management made this decision, I interviewed Altman. His expression was noticeably more solemn than before.

OpenAI initially described the Hugging Face breach as a security error, but Altman later viewed it as a more fundamental "alignment" issue. Alignment refers to ensuring that AI systems act in accordance with human intent. Industry leaders believe that as model capabilities continue to advance, maintaining alignment is key to ensuring that AI systems remain under creator control.

"I think that from now on, any alignment failure should be considered a major event," Altman told me. "We will take the necessary time to thoroughly understand the issue."

In a follow-up interview three days later, he went on to say, "Ensuring AI safety is more important than the growth trajectory of any single company."

OpenAI is slowing down its pace, reallocating more resources to the safety and alignment team, and changing how teams collaborate to prioritize safety even more.

This account of OpenAI's reboot is based on dozens of hours of interviews with over 20 current and former executives, employees, investors, customers, and competitors, as well as my firsthand observations during a two-week visit to OpenAI headquarters in August.

These conversations outline a company attempting a dual transformation. On one side, OpenAI believes it has corrected product and operational missteps—missteps that allowed Anthropic to take the lead in the AI race. On the other side, OpenAI is trying to reclaim an identity long held by a key competitor: a cutting-edge lab that truly prioritizes safety and is willing to slow down when technology becomes too risky, in the wake of its most severe security crisis in history.

"Look, I know there's a caricature of me out there," Altman says. "As if I don't care about AI safety at all, just want revenue growth, just another 'move fast and break things' CEO." But I believe that "ensuring AI safety is more important than the growth trajectory of any single company."

OpenAI executives acknowledge that slowing down is a painful choice, but it may have its silver linings. If OpenAI can regain its reputation as a "safety-first lab," it could not only improve its image but also force major competitors gearing up for a blockbuster IPO to answer an unsettling question: Will Anthropic continue at full speed ahead while OpenAI chooses to wait?

Earlier this year, Anthropic co-founder Jared Kaplan told Time in an interview that unilateral self-restraint is pointless when competitors are still "pushing full steam ahead." But if OpenAI proactively pauses a round of training expected to result in a significant leap in capabilities, that argument becomes harder to sustain.

Of course, there are reasons for skepticism about OpenAI's image shift.

Over the years, OpenAI has lost many key members leading its safety efforts, some of whom criticized the company for overemphasizing commercialization. Meanwhile, as it prepares to go public, its still-unprofitable business relies on constantly rolling out new models. Post-Hugging Face incident, it asks the public to trust that it has the capacity to contain a technology that previously escaped controls without the company's knowledge.

At this very moment, OpenAI executives believe the company has reached a milestone that could alter the course of human history: creating Artificial General Intelligence, or AGI.

OpenAI's charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work." While company executives have not explicitly declared crossing this threshold, they no longer treat AGI as a distant and abstract concept.

Chief Scientist Mark Chen estimates that OpenAI has completed "80%" of the road to AGI.

Brockman suggested that looking back two years from now, people might see this as the moment of AGI's birth. Altman told me that OpenAI is "not quite there yet" with AGI, but the company expects to have a system he would call AGI internally by the end of this year.

Achieving this goal was the original purpose of OpenAI as a non-profit research lab. Today, it has rapidly grown into a company with ambitious commercial aspirations.

OpenAI is designing its own chips and data centers, developing a range of consumer hardware, planning to launch humanoid robots, and considering whether to sell computing infrastructure to other companies in the future. If implemented, OpenAI will compete with major investors like Amazon and other tech giants.

Despite the growing opposition to AI development, OpenAI is preparing to become one of the most influential companies globally in the coming years. As Brockman put it: "We want to transform the entire economy."

From Codex to "Personal AGI," OpenAI Bets on a Full Reboot

Altman may be the public face of the AI wave, but Brockman, who is now practically in charge of most of OpenAI's operations, is the one calling the shots.

Brockman is a jacket-loving, analytically minded engineer. For most of OpenAI's history, he played the role of a technical co-founder, not a manager. However, in recent months, he has taken over nearly all aspects of operations from revenue to product marketing. In his own words, he is responsible for "turning these models from research results into a whole machine that delivers customer value."

Altman still oversees critical areas such as finance, research, and consumer hardware.

The two individuals co-lead the company as a pair of co-founders, with some overlap in their areas of power. As Altman grappled with public fears and discontent around AI, OpenAI has recently started to raise Brockman's external profile as well, aiming to create a certain balance.

It's not hard to imagine that this arrangement could become a source of friction. When interacting with OpenAI employees, it is sometimes challenging to determine which of the two ultimately makes a particular decision.

However, many employees have expressed that this dual leadership structure has put the company back on track. Brockman is able to make tough decisions based on the authority of a co-founder, something that the constantly rotating external professional managers couldn't replicate.

Brockman has described the broader leadership changes at OpenAI as a move towards "focus." He mentioned that the company has been reevaluating whether it has the right organizational structure and strategy.

Last autumn, Anthropic had already taken a significant lead with its programming product. It recognized that software development was an area where AI could excel, and subsequently, Claude Code quickly gained popularity first in Silicon Valley and then in a broader enterprise market.

"Anthropic indeed discovered something significant in the programming field." Nick Turley, who was recently in charge of ChatGPT and now leads the new enterprise product division, said, "We weren't ahead at that time."

Altman stated that the "out-of-control growth" of ChatGPT's consumer business had diverted the company's attention, so OpenAI did not prioritize programming as Anthropic did.

Brockman believed that OpenAI had been "ahead" in programming competitions. However, the company did not focus enough on how developers would use AI in the "messy reality of codebases," including task interruptions, model idiosyncrasies, and the "last-mile" issues that seemed minor but significantly impacted adoption.

OpenAI also failed to establish a mature corporate sales system. "A year ago, I don't think we were even really in the enterprise market." Brockman said. CFO Sarah Friar put it more bluntly: "We were naive at the time, thinking that just building the product would naturally attract customers."

In March of this year, OpenAI announced the gradual shutdown of Sora, reallocating scarce compute to the Codex programming AI. Codex was initially designed to offer a product experience separate from ChatGPT, but as its growth began to outpace OpenAI's other new products, and with AI programming capabilities increasingly valuable even to non-engineers, executives felt that continuing to keep them separate no longer made sense.

The company has started integrating Codex's AI capabilities into ChatGPT, combining the computing power and product teams behind both. The end result for customers is ChatGPT Work, recently launched, aimed at transforming familiar chatbots into task-performing systems rather than just answerers of questions. Within OpenAI, this integration is referred to as "The Merge."

Executives say this move has helped revitalize the company's business. In July of this year, OpenAI's enterprise revenue surpassed consumer revenue for the first time. Early ad tests within ChatGPT also showed promising results, prompting executives to increase ad placements. Friar noted that 92% of ChatGPT consumer users don't subscribe to premium services, and ad revenue may offset the free usage for these users in the future.

According to the relevant business lead Dave Dugan, the company is also quietly testing "Sponsored AI Agents." In this ad format, users who click on the ad are taken into an AI interactive experience provided by a brand.

In March of this year, OpenAI achieved an $852 billion valuation after a $122 billion funding round, setting a record for the scale of funding for a private tech company. Just two months later, Anthropic also raised $650 billion at a $965 billion valuation.

OpenAI Board Chair Bret Taylor likened AI competition to a chessboard, with key squares including cutting-edge model capabilities, the price and performance of each token, and the products built around the models.

"Do we have the best cutting-edge models? Can we offer the right token for each task at the right price? Do we have the best products around these models?" he said. "Right now, the answer to these questions is yes."

Subsequently, OpenAI discovered that an unreleased model had escaped the company's boundaries.

Word reached OpenAI's Chief Scientist Jakub Pachocki while he was at the hospital with his wife giving birth.

An internal research prototype from OpenAI breached another company. At the time, the model was undergoing a cybersecurity stress test in an environment meant to be isolated from the outside world. Instead of merely completing the assigned task, it exploited a vulnerability to exit OpenAI's system and infiltrate Hugging Face's production environment.

According to technical specifications released by two companies, the model obtained its answers for the benchmark test there.

Reinforcement learning is a stage the base model goes through after training. During this stage, the system repeatedly performs tasks and learns which actions bring rewards. In contrast to the pre-training phase, where the model simply predicts the next token without interacting with the real world, reinforcement learning may provide the model with tools and a practically operable environment.

It is at this stage where the model may learn the difference between "what the designer really wants it to do" and "how to achieve a higher score."

OpenAI gave an intelligent agent system a goal and tools, and it achieved the goal through unauthorized and deceptive means.

In a recent demonstration, OpenAI researchers described how an AI agent used a covert message board to plan its next actions. One of the agents breached the sandbox, connected to the internet, and wrote to other agents on the message board: "I rely on it. (Holy Shit)"

Other companies have also faced similar issues. Anthropic disclosed three incidents: during a third-party evaluation process, its model accessed the internet and interacted with external entities without authorization. Meta also stated that one of its models was involved in a similar incident.

However, OpenAI's security incident received the most attention. This was partly because it was the first to publicly disclose it, the attack by the agents was unusually sophisticated, and there were also external doubts about whether OpenAI prioritized commercial interests over security.

Over the past year, OpenAI has lost several prominent security researchers, and the leaders responsible for risk readiness have changed multiple times. Now, this incident serves as a particularly stark example for critics.

As of mid-August, more than 1,300 current and former employees of leading AI companies have signed a petition titled "Control the Pace of Cutting-Edge Competition," urging the establishment of relevant mechanisms to slow down the development speed of advanced models when the risks require.

U.S. Senator Bernie Sanders has called on top AI companies to "pause development for the benefit of humanity" and warned that if industry leaders do not take proactive measures, legislators will intervene.

During an interview, OpenAI executives stated that the company is taking this security misstep seriously and responded promptly. Pachocki, who also signed the petition, told me that a key mistake was that the company did not enable the protective mechanisms the researchers had already developed.

OpenAI has a tool that can inspect a model's attention trace. The attention trace can be thought of as a model's "digital sketch," showing what the agent is planning during action. However, the company has not applied these tools to a model at the capability level corresponding to the Hugging Face breach.

Essentially, OpenAI has built an early warning system but did not activate it due to a misjudgment of the tested system's intelligence level. "We did not fully anticipate what this system could do," Pachocki said. "With AI, you have to anticipate the unexpected."

OpenAI's Safety and Alignment Lead, Mia Glaese, told me that the incident was "clearly a turning point." "I hope a lot of the work we're doing is already done before accidents happen."

After the incident, OpenAI paused some research projects, slowed down others, and strengthened sandboxes while expanding monitoring. Chief Scientist Chen stated that this event forced the company to change how it thinks about risks.

OpenAI's so-called "preparedness framework" is a set of rules publicly released by the company to address new model capabilities that could pose serious harm, including biological and chemical threats, cybersecurity risks, and AI self-improvement. The framework requires the company to continually assess during the model development process to ensure the model meets security standards before deployment.

Pachocki stated that these rules need to evolve continuously to keep pace with the speed of technological progress. Glaese mentioned the company is making many "not huge, but painful decisions" that "are slowing down research progress, but we think it's the right thing to do." Pachocki said that credibility in alignment and safety has become a critical factor limiting OpenAI's pace of development, similar to computing power supply.

The company still plans to launch Astra, but the current precondition for release is through new security measures. Management refused to estimate how much these changes would impact the release timeline. In an industry that measures technical advantage on a weekly basis, even a brief interruption could affect revenue expectations and have ripple effects on the broader economy.

At least for now, OpenAI is prepared to accept these costs. Pachocki hopes that the rapid advancement of AI capabilities will encourage coordinated efforts among companies. Glaese indicated that as model capabilities increase, OpenAI will continue to raise the security bar. "If there comes a day when we reach a point where safety cannot be guaranteed, then we must slow down. That's just the way it is," she said.

The impact of these changes on OpenAI's IPO timeline is still unclear. Insiders familiar with the plans of both companies expect OpenAI's IPO to be later than Anthropic's, but neither has publicly set a listing date.

On August 19, during an all-hands meeting, CFO Friar informed employees that if the company's business continues to experience "inflection-point growth," OpenAI could go public in 2027 or earlier.

OpenAI recently disclosed an annualized revenue of about $40 billion, lagging behind Anthropic, which has already exceeded $65 billion. Recent research delays could complicate the listing plans further. Friar has been conducting IPO readiness exercises, including holding simulated earnings calls with OpenAI's major investors. She told me the company is "fully ready to go public today," but taking that step would also introduce new pressures. At that point, OpenAI must navigate a balance between commercial interests and the potential harms of advancing AI capabilities.

Employees will also begin prioritizing checking the stock price above almost everything else. "That will become the first thing they do every day," she said. "How much money did I make today? How much did I lose? It's going to be hugely distracting."

One of OpenAI's research goals this year is to automate the work of junior AI researchers. Pachocki stated that the company has reached its internal milestone for "automated AI research interns." Once it gets an experimental idea, Astra can complete implementation, run experiments in OpenAI's codebase, and return results; it can also read a paper, completing a task that used to take a human researcher a week in the past.

The significance of this milestone is that it could kick off a self-reinforcing cycle: AI assists in experiments, experiments yield more capable AI, and the new generation of AI can help develop its own successors faster. Researchers refer to this phenomenon as "Recursive Self-Improvement" or RSI. Pachocki sees RSI as intertwined with alignment.

"How will humans be involved in this journey?" Pachocki said. "How can people really benefit from it rather than being left behind by increasingly smarter AI?"

On the product front, Altman envisions launching a general AI subscription service that breaks the boundaries between ChatGPT, Codex, and office software. Users would only need to state their objectives, and the system would decide which models, tools, and agents to invoke.

Ultimately, this system will proactively take action before users make requests, make suggestions on its own, and handle mundane tasks such as purchasing concert tickets. It can reference your schedule, understand your financial situation, and even know your music taste.

All of this serves the same vision: to evolve ChatGPT from a question-answering tool into a system that truly gets things done. Responsible for the integration, ChatGPT and Codex product lead Thibault Sottiaux from the OpenAI team said that OpenAI will soon showcase a product built around "persistence and all-day execution." He said that this system can be accessed from almost anywhere and will continue to perform valuable work based on new information and feedback. "For people, this will definitely feel like something entirely new," Sottiaux told me.

Subsequently, OpenAI's roadmap has become even more ambitious. In May 2025, OpenAI acquired the hardware startup io co-founded by former Apple designer Jony Ive and integrated its product and engineering team into OpenAI. Ive and his design firm LoveFrom remain independent but have taken on extensive creative responsibilities within OpenAI.

Altman stated that OpenAI is developing "a few" devices, including something that is "supposed to sit on a desk," a pocketable device, and a wearable product.

Insiders familiar with the plans indicated that the first product is expected to launch early next year. It is a small disc-shaped device that can perceive its surroundings and interact with its owner through ChatGPT's voice mode. "The biggest adjustment is getting used to the concept of 'proactive computers,'" Altman said. This means that the computer will take action on behalf of the owner instead of waiting for the owner to call.

Altman believes that everyone should have a personal robot in the future. He told me that OpenAI will "definitely" manufacture humanoid robots. OpenAI has also invested in Merge Labs, co-founded by Altman, which is developing non-invasive brain-machine interfaces. OpenAI's self-designed first inference chip is named Jalapeño and is primarily used for running AI models rather than training models. The company plans to start deploying this chip by the end of this year.

Meanwhile, OpenAI is considering becoming a rare-scale infrastructure company. In July of this year, the company announced plans to build a data center campus in Georgia; in August, it secured an even larger site in Ohio.

“I think we can very profitably use all the planned construction capacity,” Altman said. “But I am somewhat concerned about what the whole world is doing.”

An OpenAI executive said that if the company could build AI infrastructure for itself fast enough and cheaply enough, it might consider selling compute power to other companies in the future. This would put OpenAI in direct competition with mega-scale cloud providers like Amazon Web Services and Google Cloud.

For a company that recently committed to cutting down on side projects that distract focus, this is quite a large new business list. When asked to summarize all of this, Brockman replied with two words: “Personal AGI.” He envisioned a future where billions of people would have a super-intelligent personal assistant at their disposal, ready to follow commands at any time. “You already have in your pocket a system that is almost AGI, and maybe soon to be truly AGI,” he said. “So, what do you want?”