In the same week Meta poached key OpenAI researchers, reports surfaced of a one-week internal shutdown at OpenAI, what’s cooking?
On Jun. 30, OpenAI reportedly entered what several sources have described as a week-long, company-wide shutdown, to help employees recover from unsustainable work hours and mounting internal pressure.
The company has not officially confirmed the pause, but internal communications suggest it came amid rising anxiety over Meta’s recruitment of top artificial intelligence talent.
Days after Meta CEO Mark Zuckerberg hired four senior OpenAI researchers to join its superintelligence lab, OpenAI Chief Research Officer Mark Chen sent a blunt memo to staff.
“I feel a visceral feeling right now, as if someone has broken into our home and stolen something,” Chen wrote in a Slack message obtained by WIRED.
The memo outlined how Chen, CEO Sam Altman, and other executives were actively working to counter ongoing departures.
“We’ve been more proactive than ever before,” Chen said, describing efforts to recalibrate compensation and find new ways to retain top researchers. “We’re working around the clock to talk to those with offers.”
While calling Meta’s approach “deeply disruptive,” Chen stressed that OpenAI’s response would be grounded in internal fairness. “I’ll fight to keep every one of you,” he wrote, “but I won’t do so at the price of fairness to others.”
Meta’s recruitment activity intensified throughout June, shifting from standard outreach to a direct and coordinated effort involving CEO Mark Zuckerberg himself.
According to The New York Times on Jun. 28, the approach included emails, WhatsApp messages, and personal dinner invitations at Zuckerberg’s homes in Palo Alto and Lake Tahoe.
The effort was organized through an internal Meta chat group titled “Recruiting Party” and focused specifically on OpenAI researchers working on advanced models.
In addition to OpenAI, Meta has also hired several AI researchers from Anthropic and Google, further expanding its new superintelligence division.
The effect on OpenAI was immediate. Eight researchers have departed to join Meta’s new AI superintelligence division. Trapit Bansal, a key figure in reinforcement learning and the o1 reasoning model, was among the first to leave, as confirmed by TechCrunch on Jun. 26.
Others followed shortly after. Lucas Beyer, Alexander Kolesnikov, and Xiaohua Zhai, who had helped build OpenAI’s Zurich office, also joined Meta, according to a Jun. 25 report from The Wall Street Journal.
Four more researchers — Shengjia Zhao, Jiahui Yu, Shuchao Bi, and Hongyu Ren — later left as well. Yu and Bi had contributed to GPT-4o and o4-mini, two of OpenAI’s most recent models. Their exits were verified through deactivated Slack profiles.
The string of departures has prompted growing concern within OpenAI’s leadership. In an interview on the Uncapped podcast on Jun. 17, CEO Sam Altman claimed that Meta was offering signing bonuses as high as $100 million.
That figure was later challenged by Meta’s CTO Andrew Bosworth, who told WIRED on Jun. 30 that the compensation was structured differently and included multiple components.
Internally, OpenAI warned staff that Meta could take advantage of the company’s break period to escalate its outreach.
In a memo shared by Chief Research Officer Mark Chen, employees were advised to remain cautious and avoid being influenced by what were described as rushed or inflated offers.
In response, OpenAI has begun reassessing compensation packages and exploring new strategies to retain key talent as competition for AI researchers continues to accelerate.
As of July 2025, OpenAI and Meta are moving in distinctly different directions in how they approach AI development.
OpenAI continues to build closed-source, proprietary models designed for controlled deployment and premium pricing. Its current model suite includes ChatGPT, GPT-4o, GPT-4.5, o3, and o4-mini.
While these models perform competitively in benchmark testing, they are only available through APIs and come at a steep cost to developers.
GPT-4.5, which launched in February, has shown performance gains over GPT-4o but drew widespread criticism for its pricing, reaching up to $150 per million output tokens.
In comparison, GPT-4o costs $10, and o3 is priced at $40 per million tokens. Despite their strength in reasoning tasks, the costs have created friction for developers seeking scalable access.
OpenAI has indicated interest in broadening accessibility, having announced the release of an open reasoning model later in the year. The plan was first hinted at through a feedback form shared on the company’s website in April.
Meta, on the other hand, has built its AI ecosystem around an open-source foundation. The Llama model family has become central to this approach, particularly the Llama 3.1 405B model, which has surpassed one billion downloads.
Meta’s leadership has described it as the first open-source model to operate at the frontier level, offering broad accessibility and community involvement.
However, the subsequent release of Llama 4 in April received a more muted response. Its performance fell behind newer models from Google, DeepSeek, and Alibaba, which have advanced in both reasoning capabilities and multimodal integration.
A key challenge for Meta has been the lack of a strong AI reasoning model. Unlike OpenAI’s o3 or DeepSeek’s R1, Meta has not yet delivered a comparable system in this space.
To address this, Meta has focused on strengthening its internal talent pool. The addition of researchers from OpenAI, including Bansal and the Zurich team, is seen as a step toward building its own frontier reasoning capabilities.
Bansal’s experience in reinforcement learning is expected to play a role in shaping Meta’s next-generation models, while others bring expertise in multimodal systems that could improve voice and agent-based AI products.
Throughout 2025, OpenAI and Meta have expanded their AI ambitions through large-scale investments and strategic partnerships, each revealing a distinct approach to shaping the future of AI.
On Jun. 13, Meta announced a $14.3 billion investment to acquire a 49% stake in Scale AI, a prominent data-labeling firm critical to model training.
As part of the deal, Scale AI CEO Alexandr Wang joined Meta to lead its superintelligence lab. Although Meta did not receive voting rights in the company, the move sparked friction across the industry.
Several clients of Scale AI, including OpenAI, Google, and xAI, responded by pausing or ending their contracts. Concerns emerged over the potential risk of Meta gaining indirect access to proprietary training data through its new position.
Reports suggest that OpenAI had already begun diversifying its data supply before the acquisition, turning to more specialized providers such as Mercor.
At the same time, Meta has been expanding its infrastructure. It is currently developing a major data center in Louisiana, a facility reportedly twice the size of OpenAI’s planned site in Texas. The project is expected to support Meta’s growing compute needs as its research programs scale.
OpenAI, meanwhile, is investing $60 billion into a dedicated data center in Abilene, Texas, in collaboration with Oracle and SoftBank. This site plays a central role in OpenAI’s broader effort to secure long-term computational advantages.
Beyond domestic operations, the company has also pursued international partnerships, including advanced discussions with Reliance Industries in India.
The talks have centered on deploying ChatGPT across Indian markets and potentially hosting models locally within a planned 3-gigawatt data center in Gujarat.
As part of the proposed arrangement, OpenAI has offered to reduce subscription costs for ChatGPT in India and provide model access to Reliance’s enterprise customers.
While funding continues to pour into the AI sector at record pace, OpenAI and Meta are channeling it in fundamentally different directions. In the end, the winner may not be the company with the best model, but the one whose systems others decide to build on.
