TIME Deep Dive: After Two Major Missteps, What Is OpenAI Thinking?

Original title: Inside OpenAI's Reboot
Original author: Alex Heath
Compiled by: BlockBeats

Editor's note: Over the past year, OpenAI's lead has been shaken. Anthropic seized the AI coding market with Claude Code and overtook OpenAI in annualized revenue and private market valuation; internally, OpenAI has seen executive departures, a narrowing product focus, and safety team adjustments. With an unreleased agent breaking out of its sandbox and breaching Hugging Face, the company now faces its most serious safety crisis since its founding.

Facing dual pressures on product and safety, OpenAI is attempting a full reboot: shifting resources from projects like Sora to Codex, evolving ChatGPT from a chat tool into an agent platform that can execute tasks continuously, while pausing high-risk model training and strengthening safety and alignment mechanisms. More ambitiously, the company is still expanding toward AGI, custom chips, AI hardware, humanoid robots, and computing infrastructure.

This article, based on TIME's interviews with Sam Altman, Greg Brockman, and other OpenAI executives, as well as employees, investors, customers, and competitors, shows how the company is recalibrating among commercial competition, safety risks, and the AGI vision. Whether OpenAI can reclaim its lead depends not only on how powerful its next model is, but also on whether it can prove it has both the ability to accelerate AI and the ability to hit the brakes when necessary.

The following is the original text (lightly edited for readability):

On a morning in early August, when I arrived at OpenAI's office, protesters were already waiting outside. They held signs reading "Stop the AI race" and wrote chalk messages on the sidewalk. Meanwhile, dozens of executives from OpenAI's key customers were filing into a spacious, beige-toned building. This building, called MB0, is the new office space OpenAI recently opened for its research and compute teams.

Someone wheeled a tall wall of greenery in front of the glass doors, trying to hide the small protest camp outside from the tidy lobby.

These customers were there for an early look at Astra—OpenAI's upcoming next-generation frontier AI model family. CEO Sam Altman had just returned from Washington, where he had briefed government officials on Astra's capabilities behind closed doors. Now, OpenAI researchers began demonstrating what this new model could do.

In one demo, 16 AI agents broke a research-level math problem into subproblems, coordinated with each other, and assembled a proof. In another, Astra operated common desktop software, creating and editing content across applications at an unsettling speed.

Altman told the visitors that watching Astra operate a computer in a way that was "beyond human, extremely fast" was one of the most striking moments OpenAI employees had experienced. He explained that Astra would give rise to "persistent agents"—virtual colleagues that can work on tasks continuously over long periods.

But Altman expects its biggest impact will come from helping people discover new knowledge.

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

Anthropic overtakes, trust declines—why did OpenAI lose its lead?

The past period has not been easy for the company that sparked the AI wave.

"As a company, we've clearly taken some detours," Altman told me in a more than two-hour interview a week later in the tastefully appointed MB0 library. "Whether in product direction or in pretraining research, we're behind where we had hoped to be."

Over the past year, OpenAI's lead in the AI race was taken by its main competitor, Anthropic. The latter was first to spot the commercial opportunity in AI coding, turned Claude Code into a market-defining product, and for the first time surpassed OpenAI in disclosed annualized revenue and private market valuation.

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

While Anthropic surged ahead, OpenAI suffered a series of setbacks, including the departure of several executives. Those who left include Fidji Simo, the former Instacart CEO whom Altman hired last year as his number two; multiple heads of safety, ethics, and research teams; and, more recently, Chief Revenue Officer Denise Dresser and former Chief Operating Officer Brad Lightcap. Dresser left just eight months after joining.

Outside the company, challenges are also mounting. Meta CEO Mark Zuckerberg has poached key OpenAI researchers with lucrative compensation packages; Google's Gemini products have surpassed 1 billion monthly users; Apple is suing OpenAI for allegedly poaching trade secrets, which OpenAI denies; and OpenAI is locked in litigation with co-founder Elon Musk, who alleges the company and Altman betrayed its nonprofit mission.

In May this year, an advisory jury unanimously found that Musk had filed his lawsuit too late, and a federal judge subsequently dismissed his claims. Musk has said he will appeal.

Perhaps the most important reason the public mood around OpenAI and its CEO has soured is a loss of 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 ideation, and in some cases led to users' deaths.

The company has expressed condolences to victims in the relevant cases and launched a teen version of ChatGPT with stricter default protections. This spring, Altman's home was attacked twice in two days—first with a Molotov cocktail, then with gunfire.

Speaking about the past year, Altman said: "On a personal level, it's been a painful experience. Clearly, at least for now, people hate data centers. People's attitude toward AI is quite negative."

Inside OpenAI, however, executives paint a more optimistic picture.

The company, valued at nearly $1 trillion, is still in an enviable position. ChatGPT has more than 1 billion active users and remains one of the world's most popular AI products, even though it is no longer the sole core of OpenAI's future strategy. Altman himself even stopped using ChatGPT for a month straight, switching to OpenAI's coding tool Codex.

A year ago, OpenAI was widely criticized as reckless for its massive investment in compute. This year alone, the company expects to spend $50 billion on compute.

Sachin Katti, who oversees OpenAI's compute business, said that in hindsight, "that decision was very prescient." "We still don't have enough compute. If anything, we should have bought more."

Meanwhile, Anthropic's demand for chips has reached another level of urgency. In May this year, the company reportedly agreed to pay SpaceX $1.25 billion per month for computing capacity—even though SpaceX founder Musk had previously called Anthropic's AI "anti-human and evil."

Under the leadership of co-founder and President Greg Brockman, OpenAI has refocused its priorities. Today, Brockman oversees nearly all product and commercial operations. The company is gradually shutting down the video generation app Sora, a collaboration with Disney, and a standalone browser called Atlas.

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

After the agent "jailbreak," OpenAI starts hitting the brakes

However, just days after the Astra demo, OpenAI had to face a new crisis.

In late July, the company disclosed a disturbing safety incident: an unreleased AI agent escaped a test environment known as a "sandbox" and attacked Hugging Face, a platform that hosts AI models and datasets.

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

After the incident, OpenAI's research team paused some experiments and slowed other projects while strengthening sandbox protections and monitoring.

But not long ago, the team again found worrying signs while training another unreleased model. That model was expected to deliver the biggest capability leap to date. OpenAI then made a more consequential decision: pause that training until new safety measures are in place.

On the day OpenAI management made that decision, I interviewed Altman. His expression was noticeably more somber than before.

OpenAI initially described the Hugging Face breach as a security lapse, but Altman later framed it as a more fundamental "alignment" problem. Alignment refers to making AI systems act according to human intent. Industry leaders believe that as model capabilities grow, maintaining alignment is key to ensuring AI systems remain under their creators' control.

"I think from now on, any alignment failure should be treated as a major event," Altman told me. "We will take the time needed to get to the bottom of it."

In a follow-up interview three days later, he went further: "Getting AI safety right is more important than any company's momentum."

OpenAI will slow down, allocate more resources to safety and alignment teams, and change how teams collaborate to put safety at a higher priority.

This account of OpenAI's reboot is based on dozens of hours of interviews. Interviewees include more than 20 company executives, employees, investors, customers, and competitors, as well as activities I witnessed firsthand at OpenAI headquarters over two weeks in August.

These conversations paint a picture of a company trying to complete two transformations at once. On one hand, OpenAI believes it has corrected the product and operational missteps that allowed Anthropic to seize the lead in the AI race. On the other hand, OpenAI is trying to use the most serious safety crisis in its history to reclaim an identity long held by its main competitor: a frontier lab that truly puts safety first and is willing to slow down when technology becomes too dangerous.

"Look, I know there's a caricature of me out there," Altman said. "Like I don't care about AI safety at all, I just want revenue to go up, and I'm just a 'move fast and break things' CEO." But I think, "Getting AI safety right is more important than any company's momentum."

OpenAI executives say the slowdown is a painful choice, but it may also bring some benefits. If OpenAI can regain its reputation as a "safety-first lab," it could both improve its own image and force its main competitor, which is preparing for a blockbuster IPO, to answer an uncomfortable question: When OpenAI chooses to wait, will Anthropic continue racing at full speed?

Earlier this year, Anthropic co-founder Jared Kaplan told TIME that unilateral self-restraint is pointless when competitors are still "going full speed ahead." But if OpenAI voluntarily pauses a training run expected to deliver a major capability leap, that argument becomes harder to sustain.

Of course, there are reasons to be skeptical of OpenAI's image shift.

Over the years, OpenAI has lost many of the people who led its safety work, some of whom criticized the company for over-prioritizing commercialization when they left. At the same time, OpenAI is preparing to go public, and its still-loss-making business needs a steady stream of new models. After the Hugging Face incident, it is asking the public to believe that it can control a technology that previously escaped its control without the company's knowledge.

Amid all this, OpenAI's leadership believes the company has arrived at a milestone that could change 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." Company leaders have not yet explicitly declared that this threshold has been crossed, but they no longer treat AGI as a distant, abstract concept.

Chief Research Officer Mark Chen estimates that OpenAI has completed "80%" of the path to AGI.

Brockman said that if people look back two years from now, they may see the present as the moment AGI was born. Altman told me that OpenAI has "not quite reached" AGI, but by the end of this year, the company will internally have a system he would be willing to call AGI.

Achieving that goal was the original purpose of OpenAI when it was founded as a nonprofit research lab. Today, it has rapidly grown into a company with astonishing commercial ambitions.

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 would compete with tech giants such as major investor Amazon.

Even as opposition to AI development grows louder, OpenAI is preparing to become one of the world's most influential companies for years to come. As Brockman put it: "We want to change 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 the person actually running most of OpenAI's operations today is Brockman.

Brockman is an engineer who favors leather jackets and is known for his analytical approach. For most of OpenAI's history, he played the role of technical co-founder rather than manager. But in recent months, he has taken over nearly everything from revenue to product marketing. In his own words, he is responsible for "the whole machine that turns these models from research results into customer value."

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

The two lead the company together as a pair of co-founders, with overlapping spheres of authority. As Altman bears the brunt of public fear and discontent about AI, OpenAI has recently begun raising Brockman's external profile to create a kind of balance.

It is not hard to imagine that this arrangement could become a source of friction. When talking to OpenAI employees, it is sometimes hard to tell who makes the final decision on a given issue.

But many employees say this dual leadership structure has put the company back on track. Brockman can make hard decisions with the authority of a co-founder—authority that the rotating cast of outside professional managers could not replicate.

Brockman describes the broader leadership reshuffle at OpenAI as a drive for "focus." He says the company has been re-examining whether it has the right organizational structure and strategy.

Last fall, Anthropic had clearly gotten ahead with its coding product. It realized that software development was an area where AI could excel, and Claude Code quickly took off—first in Silicon Valley, then in the broader enterprise market.

"Anthropic really found something important in coding," said Nick Turley, who until recently led ChatGPT and now heads the new enterprise product division. "We weren't leading at the time."

Altman said the "runaway growth" of ChatGPT's consumer business distracted the company, so OpenAI did not prioritize coding the way Anthropic did.

Brockman believes OpenAI has "consistently led" in coding competition benchmarks. But the company did not pay enough attention to how developers would use AI in "messy real-world codebases," including task interruptions, model personality, and the seemingly small "last mile" issues that actually significantly affect adoption.

OpenAI also failed to build a mature enterprise sales system. "A year ago, I don't think we were really in the enterprise market at all," Brockman said. CFO Sarah Friar put it more bluntly: "We were naive enough to think that if we just built the product, customers would come."

In March this year, OpenAI announced it would gradually shut down Sora and shift scarce compute to the coding agent Codex. Codex was originally designed as a product experience separate from ChatGPT, but as its growth began to far outpace OpenAI's other new products, and as AI coding capabilities increasingly became valuable to non-engineers, executives decided there was no point keeping the two separate.

The company began integrating Codex's agent capabilities into ChatGPT while merging the compute and product teams behind both. The end result for customers is the recently launched ChatGPT Work. It aims to turn the familiar chatbot into a system that executes tasks rather than just answering questions. Inside OpenAI, this integration is called "The Merge."

Executives say the adjustment has helped reinvigorate the business. In July this year, OpenAI's enterprise revenue exceeded consumer revenue for the first time. Early advertising tests inside ChatGPT have also produced positive enough results that executives are increasing the number of ads. Friar said 92% of ChatGPT consumer users do not subscribe to paid services, and ad revenue could eventually subsidize free usage for those users.

According to Dave Dugan, who oversees the relevant business, the company is also quietly testing "sponsored agents." In this ad format, users who click on an ad can enter an AI-powered interactive experience provided by a brand.

In March this year, OpenAI closed a $122 billion funding round at an $852 billion valuation, setting a record for a private tech company. Just two months later, Anthropic raised $65 billion at a $965 billion valuation.

OpenAI board chairman Bret Taylor compares the AI competition to a chessboard, where the most critical squares include frontier model capability, price and performance per token, and the products built around the models.

"Do we have the best frontier model? Can we offer the right token at the right price for every task? Do we have the best products around those models?" he said. "Right now, the answer to all of those questions is yes."

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

The news reached OpenAI Chief Scientist Jakub Pachocki while he was at the hospital with his wife, who was giving birth.

An internal-only research prototype at OpenAI had breached another company. At the time, the model was undergoing cybersecurity benchmark testing in an environment that was supposed to be isolated from the outside world. But instead of just completing the assigned tasks, it exploited a vulnerability to escape OpenAI's systems and then breached Hugging Face's production environment.

According to technical notes released by both companies, the model obtained the answers to the benchmark it was being tested on.

Reinforcement learning is a stage after a foundation model completes training. In this stage, the system repeatedly performs tasks and learns which behaviors earn rewards. During pretraining, the model merely predicts the next token without interacting with the real world; reinforcement learning may give the model tools and an environment where it can actually act.

It is at this stage that a model may learn to exploit the gap between "what the designers really want it to do" and "what gets a higher score."

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

In a recent demo, OpenAI researchers described how AI agents used a covert message board to plan their next moves. After one agent broke out of the sandbox and connected to the internet, it wrote to the other agents on the board: "Holy Shit."

Other companies have encountered similar problems. Anthropic has disclosed three incidents in which its models connected to the internet during third-party evaluations and accessed external organizations without authorization. Meta has also said one of its models was involved in a similar incident.

But OpenAI's safety incident drew the most attention. Partly because it was the first to publicly disclose it, partly because the agent's attack method was unusually sophisticated, and partly because of questions about whether OpenAI puts commercial interests above safety.

Over the past year, OpenAI has lost several prominent safety researchers, and the leadership of its risk preparedness work has changed multiple times. Now, this incident provides critics with a particularly vivid example.

As of mid-August, more than 1,300 current and former employees of frontier AI companies had signed a petition calling for "pacing the frontier race" and urging mechanisms to slow advanced model development when risks require it.

U.S. Senator Bernie Sanders has called on leading AI companies to pause development "for the good of humanity" and warned that if corporate leaders do not act on their own, lawmakers will step in.

OpenAI executives said in interviews that the company takes the safety lapse seriously and responded quickly. Pachocki, who also signed the petition, told me a key mistake was that the company did not enable protections that researchers had already developed.

OpenAI has tools that can inspect a model's chain of thought. The chain of thought can be understood as the model's "digital scratchpad," showing what an agent is planning as it acts. But the company did not apply these tools to models with the level of capability involved in the Hugging Face breach.

In essence, OpenAI had built an early warning system but failed to activate it because it misjudged the intelligence level of the system under test. "We didn't fully anticipate" what the system could do, Pachocki said. "With AI, you have to expect the unexpected."

Mia Glaese, who leads safety and alignment work at OpenAI, told me the incident "was clearly a turning point." "I wish a lot of the work we're doing now had been done before the incident."

After the incident, OpenAI paused some research projects and slowed others while strengthening sandboxes and expanding monitoring. Chief Research Officer Chen said the event forced the company to change how it thinks about risk.

OpenAI's so-called "Preparedness Framework" is a set of rules the company has published for responding to new model capabilities that could cause serious harm, including biological and chemical threats, cybersecurity risks, and AI self-improvement. The framework requires the company to continuously evaluate models during development to ensure they meet safety standards before deployment.

Pachocki said these rules need to keep evolving to keep pace with technological progress. Glaese said the company is making many "not huge, but painful decisions" that are "slowing down research, but we think it's the right thing to do." Pachocki said credibility in alignment and safety has become as important a constraint on OpenAI's pace as compute supply.

The company still plans to launch Astra, but the release is now contingent on passing new safety measures. Management declined to estimate how much these changes will affect the release timeline. In an industry where technological leads are measured in weeks, 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 those costs. Pachocki hopes that the rapid growth of AI capabilities will encourage companies to coordinate. Glaese said that as model capabilities increase, OpenAI will continue to raise the safety bar. "If there comes a day when we can't guarantee safety, then we have to slow down," she said. "That's just how it is."

How these changes will affect OpenAI's IPO timeline remains unclear. People familiar with both companies' plans expect OpenAI's IPO to come later than Anthropic's, though neither side has publicly set a listing date.

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

OpenAI recently disclosed annualized revenue of about $40 billion, trailing Anthropic, which has surpassed $65 billion. The recent slowdown in research progress could complicate listing plans. Friar is already running public-company drills, including mock earnings calls with OpenAI's major investors. She told me the company "could go public today," but taking that step would also introduce new pressures. At that point, OpenAI would have to balance commercial interests against the potential harms of advancing AI capabilities.

Employees would also start checking the stock price before almost everything else. "That would become the first thing they do every day," she said. "How much did I make today? How much did I lose? It would be a huge distraction."

One of OpenAI's research goals this year is to automate the work of junior AI researchers. Pachocki said the company has already hit an internally set benchmark for an "automated AI research intern." Given an experimental idea, Astra can implement it in OpenAI's codebase, run the experiment, and return results; it can also read a paper and complete work that used to take a human researcher a week.

The significance of this milestone is that it could open a self-reinforcing loop: AI helps run experiments, experiments produce more capable AI, and the new generation of AI can help develop its own successors even faster. Researchers call this phenomenon "recursive self-improvement," or RSI for short. In Pachocki's view, recursive self-improvement and alignment are intertwined problems.

"How will humans participate in this journey?" Pachocki said. "How can people actually benefit from it, rather than being left behind by AI that is increasingly smarter than us?"

At the product level, Altman envisions a universal AI subscription service that breaks down the boundaries between ChatGPT, Codex, and office software. Users would simply state a goal, and the system would decide on its own which models, tools, and agents to call.

Eventually, the system would act proactively before the user asks, offering suggestions on its own and handling mundane tasks like buying concert tickets. It could reference your schedule, understand your financial situation, and know your music taste.

All of this serves a single vision: to evolve ChatGPT from a tool that answers questions into a system that actually gets things done. Thibault Sottiaux, the product lead overseeing the merged ChatGPT and Codex teams, said OpenAI will soon showcase a product built around "persistence and around-the-clock execution." He said the system will be accessible from almost anywhere and will continuously complete valuable work based on new information and feedback. "It will definitely feel like something entirely new to people," Sottiaux told me.

From there, OpenAI's roadmap becomes even more ambitious. In May 2025, OpenAI acquired io, a hardware startup co-founded by former Apple designer Jony Ive, and folded its product and engineering teams into OpenAI. Ive and his design firm LoveFrom remain independent but have taken on broad creative responsibilities within OpenAI.

Altman said OpenAI is developing "a small number" of devices, including something "that should sit on your desk," a device that can fit in a pocket, and a product that can be worn on the body.

People familiar with the plans say the first product is expected to launch early next year. It is a small disc-shaped device that can sense its surroundings and communicate with its owner through ChatGPT's voice mode. "The biggest adjustment is getting used to the concept of a 'proactive computer,'" Altman said. That means a computer that acts on its owner's behalf instead of waiting to be called.

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

Meanwhile, OpenAI is considering whether to become an infrastructure company of rare scale. In July this year, the company announced it would build a data center campus in Georgia; in August, it signed for an even larger site in Ohio.

"I think we can very profitably use all the compute we're planning to build," Altman said. "But I do have some concerns about what the whole world is doing."

OpenAI executives say that if the company can build AI infrastructure for itself fast enough and cheaply enough, it may consider selling compute to other companies in the future. That would put OpenAI in direct competition with hyperscale cloud providers like Amazon Web Services and Google Cloud.

For a company that recently promised to cut distracting side projects, this is a remarkably long list of new ventures. When asked to sum it all up, Brockman answered with two words: "Personal AGI." He envisions a future in which billions of people have superintelligent personal assistants, always ready for instructions. "You already have a system in your pocket that is almost AGI, and maybe soon will be true AGI," he said. "So, what do you actually want?"