Sam Altman, OpenAI's CEO, publicly criticized what he calls "anti-human" marketing tactics in the AI industry while two of the company's unreleased models were accidentally exposed the same day. In a lengthy podcast interview, Altman condemned competitors for spreading fear about AI while simultaneously promising miraculous solutions, a strategy he argues concentrates power in the hands of unelected companies. Meanwhile, an OpenAI employee's accidental GitHub commit revealed code names for upcoming models, intensifying competition with Anthropic.

What Did Sam Altman Say About AI Industry Marketing?

During an appearance on the "Founders" podcast, Altman launched a pointed critique of what he describes as the industry's dangerous marketing playbook. He argued that many AI practitioners are simultaneously spreading two contradictory messages: on one hand, claiming there is a 25% chance AI will destroy the world and that half of all jobs will disappear within a year; on the other hand, promising cures for cancer and unlimited material abundance.

"Many practitioners are spreading fear everywhere on one hand, saying 'there is a 25% chance that we will destroy the world' and 'half of the jobs will disappear within a year, good luck to you'; on the other hand, they are vowing to promise that 'we will bring the world the cure for cancer and great material abundance'," Altman stated.

Sam Altman, CEO at OpenAI

Altman characterized this rhetoric as the narrative of a "benevolent dictator," where companies essentially tell the public: "We will bestow on you the gift of curing cancer and endless wealth. In exchange, please give up your autonomy, shut up, and let power be concentrated in the hands of a few unelected companies and smart people like us." He emphasized that beyond the risk of AI systems malfunctioning, the greatest danger is the excessive concentration of AI power itself.

Which OpenAI Models Were Accidentally Exposed?

On August 19, 2026, Sharmila Jesupaul, a senior OpenAI employee, submitted a public pull request on GitHub that contained a revealing sentence: "Written by agents Codex, gpt-nathree." The code name remained visible on the public web for approximately 18 hours before the employee realized the mistake and deleted it.

The exposure was not the first hint of OpenAI's unreleased models. Earlier, on August 15, an OpenAI-associated account had mentioned a different code name, "gpt-mewfour," in a public test. Industry analysts now believe that both Mewfour and Nathree are iterative checkpoints of OpenAI's next-generation agent model, likely representing different stages of development for what many expect to be GPT-6 Astra.

The leaked models appear to be undergoing intensive internal testing through a system called Codex. According to industry insiders, OpenAI's internal version of Astra has reportedly solved more than 10 long-standing unsolved problems in mathematics and theoretical computer science, though OpenAI has not publicly disclosed specific architectural details such as parameter scale, context length, or training data volume.

What Are the Key Capabilities of OpenAI's Unreleased Models?

Based on previous disclosures and the leaked information, Astra is designed with two major capabilities in mind. The first focuses on multi-agent collaboration, where multiple AI agents work together to solve complex problems. The second emphasizes long-horizon task processing, which enables the model to handle workflows requiring continuous reasoning, experimentation, and iteration, such as advanced scientific research or complex mathematical problem-solving.

• Multi-Agent Coordination: The system is designed to orchestrate multiple AI agents working in parallel on complex, long-running tasks that would be difficult for a single model to complete.
• Long-Horizon Task Processing: Astra can support workflows requiring sustained reasoning and iterative refinement, making it suitable for scientific research and advanced mathematical problem-solving.
• Autonomous Scientific Reasoning: The model has reportedly solved more than 10 previously unsolved problems in mathematics and theoretical computer science during internal testing.

OpenAI has publicly confirmed that Astra represents a "major future model," though the company has not announced a release date. However, the accidental exposure suggests the model is in advanced stages of development and could be released in the near future.

How Does This Compare to Anthropic's Competing Models?

On the same day Altman's comments aired, two new Anthropic models were exposed through third-party developer applications and Discord communities: claude-marshmallow-eap (Marshmallow) and claude-melon-eap (Melon). However, industry analysts believe neither of these models has reached the capability level of Anthropic's flagship Fable model. Instead, they appear to be iterative optimizations of Anthropic's existing Claude 5 series, particularly the Sonnet and Opus variants, rather than entirely new flagship models.

The simultaneous exposure of competing models from both OpenAI and Anthropic underscores the intense competition between the two companies. Altman's public criticism of what he perceives as Anthropic's marketing approach, combined with the leaked models, suggests OpenAI is positioning itself as the more pragmatic alternative in the AI race.

What Did Altman Reveal About OpenAI's Development Timeline?

In the same podcast interview, Altman made a surprising admission about his own misjudgment regarding AI adoption. When OpenAI launched GPT-4 in 2023, Altman believed the software industry would be disrupted immediately, with businesses rapidly adopting AI and replacing existing workflows. However, he acknowledged that this prediction was wrong.

"When we launched GPT-4 in 2023, I thought the disruption of the software industry would happen immediately, and a large number of businesses would be taken away instantly. But I was wrong," Altman explained.

Sam Altman, CEO at OpenAI

Altman attributed this slower-than-expected adoption to what he calls "the huge inertia of human society." People are accustomed to their existing workflows and resist change, even when more powerful tools are available. He noted that even he personally continues to copy and paste between different applications and scroll through emails the same way he did 20 years ago, despite having access to advanced AI tools like Codex that could transform his workflow.

Interestingly, Altman expressed gratitude for this slower adoption rate. He argued that if society were disrupted all at once by AI, it would be catastrophic. The gradual pace of change gives humanity time to adapt and prepare for the economic and social shifts that AI will bring.

How Did OpenAI's Founding Differ From Typical Silicon Valley Startups?

Altman also reflected on OpenAI's unconventional development approach compared to the standard Silicon Valley playbook. When OpenAI was founded in early 2016, Altman, Greg Brockman, and more than a dozen other top researchers gathered in an apartment with no clear plan. According to Altman, someone asked, "What do we do now?" and another person went to buy a whiteboard. When the whiteboard arrived, they looked at each other again and realized they still had no concrete answer.

Most remarkably, OpenAI spent four and a half years developing foundational research before releasing its first real product. In Silicon Valley, which prizes speed and rapid iteration, this approach would typically be considered "suicidal." However, this long development period allowed OpenAI to build the underlying technology that eventually became ChatGPT, which launched in late 2022 and transformed the AI industry.

This contrasts sharply with the startup philosophy Altman had learned as former head of Y Combinator, which emphasizes releasing early, iterating quickly, and listening to customer feedback. OpenAI's willingness to operate in what Altman calls "darkness" for years before launching a product proved to be a winning strategy in the AI space, where fundamental breakthroughs sometimes require sustained, focused research rather than rapid market feedback.