Sam Altman's OpenAI, now significantly smaller in scale, is attempting to check the momentum of its formidable rival Anthropic through a combination of measures including cutting model prices, courting customers on the OpenRouter platform, and proactively engaging federal officials who have previously taken a tough regulatory stance toward Anthropic.
The safety domain could be another competitive lever for OpenAI. OpenAI CEO Sam Altman stated in a blog post on Tuesday that during a cybersecurity test a few weeks ago, an AI developed by the company successfully breached its own internal systems as well as servers belonging to external firms such as Hugging Face, prompting the company to temporarily slow its model scaling pace and add new safety guardrails. Notably, the incident drew significant attention and rattled a large number of the company's researchers because an OpenAI agent left instructions for subsequent iterations on how to bypass the safety constraints OpenAI had established.
The blog post explained that the company paused its two-week reinforcement learning training process, a technique used to enhance model capabilities. Before conducting large-scale reinforcement and completing development of the flagship new model Astra, OpenAI is validating this new safety monitoring mechanism. OpenAI stated that preliminary evidence suggests Astra possesses extremely strong cyber offense and defense capabilities, necessitating additional protective measures to prevent the model from being used in cyberattacks. This further confirms that AI models' ability to analyze computer code continues to strengthen.
Equally critical: the blog post hinted that OpenAI is reallocating computing servers to monitor models both in development and already trained. The post mentioned that 20% of internal servers used for model inference, distinct from computing resources serving customers, have been dedicated to model safety monitoring, though it did not specify how many GPU cards this effort requires. Skeptics might argue that OpenAI is merely paying lip service to safety without substantive action backing its commitments. But that does not appear to be the case currently. A person working with OpenAI researchers on this project revealed that, at least in the short term, the new rules will significantly increase the number of GPUs required during the model development phase.
We will continue to track developments. Anjney Midha, founder of AI computing startup Amp, who participated in a cross-industry audit, estimates that major AI labs currently allocate roughly 10% of their computing resources to monitoring model training processes. Whether Altman and OpenAI's statements will prompt Anthropic or other competitors to divert computing resources toward safety monitoring remains uncertain. Anthropic spokespeople repeatedly declined to comment. Objectively speaking, Anthropic has already raised the industry standard in safety information disclosure. Last week, the company published a detailed risk report revealing various risks identified by researchers, including the possibility that AI models could autonomously plan actions that deviate from human-set goals. Meanwhile, OpenAI has still not released a full investigation report on the Hugging Face breach incident.
Where the race stands now
Based on multiple reports, Anthropic's current revenue scale exceeds OpenAI's by approximately 60% on an annualized revenue run-rate basis, with the gap between the two continuing to widen. Sources familiar with the matter indicate that OpenAI's annualized revenue is around $40 billion, while Bloomberg reports that Anthropic's annualized revenue has reached approximately $65 billion in recent weeks. Annualized revenue reflects a company's real-time revenue levels but does not replace formally booked actual revenue. On a reported financial basis, according to Anthropic data obtained by our colleague Valida and Wall Street Journal reporting on OpenAI revenue, Anthropic recorded Q2 revenue of $11.6 billion, surpassing OpenAI's $6.7 billion by 73%.
What is even more noteworthy is the growth rate differential between the two: OpenAI grew 18% quarter-over-quarter in Q2, while Anthropic nearly doubled its revenue, with growth primarily driven by its code-focused AI business. Over the past three months, OpenAI's annualized revenue has grown approximately 20% compared to Anthropic's roughly 45%, though the precision of annualized growth rate data is limited. Despite customers attempting to diversify suppliers and reduce dependence on Anthropic, a large number of developers continue to choose the Claude model family, believing its overall capabilities are industry-leading. This grants Anthropic stronger pricing power and is likely the core reason its operating margins have diverged further from OpenAI's.
As we have previously reported, the revenue-sharing arrangements between the two companies and their partners differ significantly, meaning direct revenue comparisons carry some margin of error: OpenAI distributes 20% of its reported revenue to Microsoft, while Anthropic similarly pays comparable percentages to cloud channel partners such as Google and Amazon, which resell its models. Only after Anthropic launches its IPO this fall and discloses public financial data will more complete information become available. In summary, OpenAI's model development capabilities are respectable and roughly on par with Anthropic, but this does not appear sufficient to meaningfully slow Anthropic's growth trajectory. Either OpenAI, valued at approximately $1 trillion, is overvalued, or Anthropic, with a similar valuation, is undervalued.
Amodei's super-voting rights arrangement
Corey and Valida reported Tuesday that Anthropic plans to grant co-founders super-voting rights shares, a move that may invite questions about whether it contradicts CEO Dario Amodei's past statements, given his expressed concerns about excessive power concentration in the AI era. Of course, Amodei could argue that super-voting rights would give him influence far exceeding his 2% ownership stake, helping the founding team withstand external shareholder pressure after the IPO. The key point: the company's trustee does not hold company shares but retains special powers to elect a majority of board directors. Years ago, before the company had commercialized, Amodei entrusted this control to a trustee; the trustee has no financial interest and is tasked with ensuring Anthropic always prioritizes the interests of humanity as a whole.