Behind the scenes of palace intrigue: How did the founders of OpenAI and Anthropic fall out?

At an all-hands meeting of OpenAI in December 2020, the screen suddenly went black. Dario Amodei showed up with 6 core members to announce their resignation. The system permissions of all 7 people were revoked on the spot, and their computers were remotely locked by the company. Internally, this incident was codenamed "The Divorce".

How could a pair of entrepreneurial partners who had fought side by side for many years come to the point of such a "divorce"?

The first crack appeared as early as 2017.

There was an internal concept codenamed "Countries" at OpenAI, which was originally intended to pull together an international alliance to jointly develop AGI. Later, it gradually went awry, turning into a competition where the United States, China, and Russia contributed funds, and the one who paid more got priority access.

Dario warned in an internal email that doing so would only exacerbate the arms race, and in the worst case, it would be linked to human rights atrocities. OpenAI responded that it had never discussed "selling AGI". When Microsoft entered the game, this plan was shelved, but the first rift between the two had already been torn open.

What made Dario no longer trust Altman was the $1 billion investment from Microsoft.

Dario requested that the contract specify that even if OpenAI joined another institution closer to AGI in the future, Microsoft could not obstruct it. Altman verbally promised to handle it, but in the end it was not written into the contract. Years later, he said he had no memory of this matter.

But once the trust was fractured, every little thing afterward would blow wind into the gap.

During the GPT-3 test, Dario found that there was a longer list of approved clients that he knew nothing about. In 2020, Altman wanted to invite Kanye West to try GPT-3 in advance, and Jack Clark in charge of policy thought it was inappropriate and firmly stopped it. All the frictions eventually converged on a specific clause.

Dario proposed that the security team led by him should have veto power over the release of new models, which OpenAI rejected.

Leaving the company was never easy. Dario was once so pessimistic that he planned to move to DeepMind, and his colleague Jared Kaplan persuaded him that since he could not rest assured anywhere else, he might as well start his own company.

The brother and sister created a secret Slack group codenamed "BATNA", which refers to the fallback plan when negotiations break down. When they held meetings, they used plush toys as spokespersons: a lazy cat named Jinji, a pink bear named Maura, and another bear named Beary Bonds, with Dario's sister Daniela acting as the mediator to smooth things over.

The new company was originally supposed to be named "Sparrow Systems", but the domain name had been registered by someone else, so it was finally named Anthropic and officially established in 2021. Initially, it only planned to recruit 50 people. By the summer of 2023, it had 160 employees, and two years later, the number grew to thousands.

Six years later, the high-stakes bet these seven people placed back then has now become OpenAI's biggest rival.

In May 2026, Anthropic completed a $650 billion financing round, with a post-money valuation of $9.65 trillion, surpassing OpenAI for the first time. At that time, OpenAI's valuation stood at $8.52 trillion. The revenues of the two companies are very close: by the end of July, Anthropic's annualized revenue reached $650 billion, while OpenAI's was just over $400 billion.

In terms of enterprise large model spending, statistics from Menlo Ventures show a gap of 40% vs 27%, with Anthropic taking the lead. In the code segment market, Claude Code once accounted for about 54% of the share.

The head-to-head confrontation has come out into the open.

On October 7, OpenAI rolled out GPT-6 to all ChatGPT users worldwide, with the paid version named GPT-6 Sol and the free version named GPT-6 Luna. On the same day, Anthropic released Claude Haiku 5.5, which costs only $0.10 per million input tokens, about 75% cheaper than the previous generation.

OpenAI still has 1.1 billion weekly active users and 2.5 million enterprise customers, so its foundation is not weak. Anthropic is speeding ahead on the path to listing, and it is said that it will start the roadshow in November, with a target valuation even more aggressive than that of OpenAI. The two companies are becoming more and more similar in their operations, and even their troubles have begun to align.

At the India AI Summit in February 2026, Narendra Modi invited a group of AI leaders to hold hands for a group photo, but neither Altman nor Dario reached out, and their elbows barely touched.

The saying that "a mountain cannot accommodate two tigers" fits their situation perfectly.

The AI game is far from dealing the last card. Who will have the last laugh?

This article is from the WeChat official account "Gelonghui APP" (ID: hkguruclub), written by Gejila, and authorized for release by 36Kr.