A new book reveals the real reason Anthropic split from OpenAI

Why it matters: the idea reportedly came from Greg Brockman, OpenAI's co-founder and president, as part of a broader vision in which OpenAI would essentially coordinate AI systems for governments worldwide, rather than simply operating as a research lab pursuing safety-focused science.

That's a genuinely different mission than the nonprofit safety-research image OpenAI publicly projected at the time. Selling access to the most powerful AI systems ever built, to any government willing to pay for it, raised the kind of question that doesn't have an easy or comfortable answer: which governments specifically, and under exactly what conditions.

By the numbers, or rather by the words used: Amodei reportedly said selling AGI to hostile countries was close to "an act of treason," and seriously considered resigning from OpenAI over the proposal alone.

That's an unusually strong, dramatic internal reaction for what started out as a fairly ordinary, run-of-the-mill corporate strategy planning disagreement. It signals this wasn't a minor philosophical difference of opinion, it was, in Amodei's apparent view, a fundamental question about who should control the most consequential technology his team was building.

In real life imagine finding out your company's leadership was seriously discussing selling a weapon-grade technology to the highest bidder among world governments. That's roughly the scale of disagreement being described here, even if AGI itself remains unbuilt.

What's next, or rather what happened next: the disagreement didn't resolve quickly. Reporting indicates tensions escalated over the following years, through 2018 and into 2020, as Sam Altman became OpenAI's de facto leader and Brockman's role in the company's direction grew.

By the end of 2020, Amodei set conditions for staying: direct reporting to OpenAI's board, and no longer working alongside Brockman. When those terms went unmet, negotiations broke down entirely. Amodei, his sister Daniela Amodei, and roughly 10 other colleagues left to found Anthropic.

• 2016: Amodei joins OpenAI, early friction with Brockman over transparency.
• 2017: The AGI-to-governments proposal creates major internal conflict.
• 2018-2020: Altman consolidates leadership; tensions continue building.
• Late 2020: Amodei's departure conditions go unmet; Anthropic is founded.

Who's affected: this isn't a single anonymous tip. Roose's book draws on more than 150 interviews with people inside OpenAI, Anthropic, and Google DeepMind, along with unpublished internal documents and research memos. That's a substantial reporting base for a story this specific.

The book reportedly also covers new details on Sam Altman's 2023 firing and reinstatement as OpenAI's CEO, suggesting it offers a broader reexamination of the company's internal history, not just this one 2017 episode.

Background: OpenAI was founded in 2015 as a nonprofit, with a stated mission to ensure AGI "benefits all of humanity" rather than the private gain of any one person or company. That framing is exactly why a 2017 proposal to sell AGI access to governments would have struck insiders as such a sharp departure.

The organization's structure shifted dramatically not long after. In 2019, facing costs that donations alone couldn't cover, OpenAI created a "capped-profit" for-profit arm still controlled by the nonprofit, capping early investor returns at 100 times their investment while directing anything beyond that back to the nonprofit's mission. Microsoft invested $1 billion that same year, eventually taking a stake reported around 49% of the for-profit entity.

Why it matters: that same tension, mission-driven nonprofit ideals colliding with the practical need for massive capital, runs through both the 2017 government-sales proposal and OpenAI's broader restructuring. Amodei's objection wasn't an isolated personality clash.

It was one early flashpoint in a fight over what kind of company OpenAI was actually going to become, a fight still playing out publicly years later as OpenAI moved even further from its nonprofit roots toward a full Public Benefit Corporation structure.

Yes, repeatedly, even years later. At a 2026 AI summit in India, Altman and Amodei were photographed visibly avoiding a handshake, a moment widely noticed given both men have appeared together at more formal events, including the UN Security Council, where they jointly called for global AI safety rules just this September.

The catch: that contrast, cooperating publicly on policy while privately maintaining a well-documented personal rivalry, is exactly the kind of complexity a corporate press release never captures. Public statements and actual working relationships between AI's most influential leaders don't always match.

What's next, in the sense of how the story ended: Anthropic launched in January 2021, founded as a public benefit corporation, a structure legally required to weigh positive social impact alongside profit rather than profit alone. That's a direct structural echo of the exact tension Amodei's team said drove them out of OpenAI in the first place.

The company's early research agenda crystallized around Constitutional AI, a training method that uses an explicit set of written principles to guide how models critique and refine their own behavior, rather than relying purely on human feedback after the fact. It became one of Anthropic's most distinctive technical contributions, and a direct product of the safety-first philosophy the founding team said OpenAI wasn't fully committed to.

Who's affected: that distinction matters for anyone choosing between the two companies' products today. Anthropic has built its entire public identity specifically around the idea that safety and capability shouldn't have to trade off against each other, a positioning that traces directly back to this exact 2017 disagreement rather than being an abstract marketing choice dreamed up separately, later, by a branding team.

• This won't change any AI product you use today, but it explains real structural history behind why OpenAI and Anthropic exist as separate, competing companies.
• The 2017 proposal underscores how early and how seriously "who controls AGI" was being debated, years before most of the public was paying attention to AI at all.
• Company origin stories matter for understanding today's AI safety debates. Anthropic's stricter public safety stance traces back to this exact kind of internal disagreement.
• Expect more detail as the book releases fully. This Atlantic excerpt is reportedly just the first piece of a larger, deeply sourced account.