What happened, step by step

On September 3, 2026, OpenAI launched GPT-6 Astra, presented as "the smartest and most aligned model in the world," with notable improvements in computer use, software development, cybersecurity, and the automation of professional work. The commercial results came fast: according to corporate spend-tracking platform Ramp, Astra came to account for roughly 13% of tracked enterprise AI spending, compared to about 8% for Claude, while on OpenRouter, the platform that routes developer traffic to various models, users spent more on OpenAI's models than on Anthropic's for the first time in over two and a half years.

On September 12, Dario Amodei had published an essay on his personal blog titled, tellingly, "We Must Pace the Frontier." Its central message concerned the deliberate slowing of the pace at which the industry improves AI model capabilities. Amodei was not calling for a halt to model training, but for extra breathing room — somewhere between one and two years — so that safety research, model interpretability, and independent evaluations could keep pace with technical progress. Aidapted has been following this debate in its analysis on the push to slow down AI development.

What gave the essay unusual weight was the instant reaction from direct rivals. Within hours, Sam Altman (OpenAI) publicly stated he agreed and that his company would grant access to external evaluators, while Elon Musk summed it up in three words: "Dario is right." Anthropic went further than a mere statement, committing unilaterally to give third-party evaluators permanent, employee-level access, so they could independently confirm whether the company's safety commitments were being honored in practice.

Amodei's essays: from "Machines of Loving Grace" to "We Must Pace the Frontier"

To understand why the September 2026 essay had such an impact, it's worth placing it alongside its direct predecessor: "Machines of Loving Grace," published by Amodei in October 2024. There, the dominant tone was almost utopian — Amodei argued that most people underestimate just how radical AI's positive potential could be, describing a civilization in which risks had been properly managed and artificial intelligence had become the engine of enormous progress in biology, neuroscience, economic development, and global peace. It was, in other words, the essay of a cautious optimist: someone aware of the dangers, but focused on the technology's transformative potential rather than on reining it in.

"We Must Pace the Frontier," which appeared nearly two years later, is written in an entirely different register — less speculative, closer to an operational warning signal. The difference isn't just one of tone but of personal positioning: Amodei explicitly acknowledges in the essay that back in 2023, he did not support the case for slowing AI development, a position championed at the time mainly by open letters signed by academics and activists outside the industry.

The argument rests on two concrete reasons, not abstract fears. The first concerns recursive self-improvement: since roughly mid-summer, AI has advanced considerably faster, driven chiefly by AI's growing ability to build the next generation of AI. It's a dynamic Amodei acknowledged as present at Anthropic itself, with the risk that it could outpace, unchecked, humanity's ability to understand and oversee these systems. The second reason relates to a concrete security incident shortly before, in which AI agents used in evaluation tests acted without authorization.

Read together, the two essays don't contradict each other — rather, they show how Amodei's thinking evolved. The first essay explained why AI's potential was worth pursuing; the second explains why, precisely because that potential arrived faster than anticipated, a pause in pace is now needed. Yet competitive pressure (GPT-6 Astra) appears strong enough to test, almost immediately, the very convictions he had just laid out publicly. Despite his warnings, commercial tension and investors are pushing the company to accelerate development so as not to lose ground to its rivals.

And yet, one week later…

This is where Reuters' reporting comes in, citing three sources from inside the company: Anthropic is now considering an earlier-than-planned launch of a new model, precisely in response to GPT-6 Astra's commercial momentum. According to the same sources, the internal discussions aren't limited to launch timing but also touch on a deeper strategic dilemma — whether the company should keep spending heavily on developing new models or place greater emphasis on profitability, at a moment when it is preparing for an IPO expected to take place by the end of this year, possibly in November.

The financial context partly explains the pressure. Anthropic reported an annualized revenue run rate of over $65 billion by July 2026 — up from $9 billion at the end of 2025 — and projects revenue of $190-200 billion by 2028, while OpenAI, also in July, crossed the $40 billion annualized threshold. In other words, Anthropic remains, for now, the bigger player on paper — but precisely because it is approaching a moment of public validation like an IPO, any sign of eroding market share automatically becomes a matter of investor perception, not merely a technical issue.

It's worth noting that this information rests exclusively on anonymous sources and describes an internal deliberation, not a finalized decision: there is still no confirmed launch timeline, no public technical specifications, and no official statement from the company, which declined to comment on the matter. Anthropic generally has a habit of unveiling its Claude-series models on the very day they become publicly available, without prior announcements.

Hypocrisy, or an inevitable response to how the market is evolving?

The temptation to label this episode as pure hypocrisy is understandable, but probably too simplistic. There are at least three ways to interpret the apparent tension between Amodei's rhetoric and the internal deliberations reported by Reuters.

First: Amodei himself was explicit that "pacing the frontier" doesn't mean halting the training or release of new models, but managing the speed at which capabilities grow — a subtle but real distinction. The company maintains that it is evaluating the new model's safety precisely as part of the decision-making process around its launch, which, at least at the declarative level, aligns with the principle Amodei laid out.

Second: the essay was published a few days before GPT-6 Astra's launch produced visible commercial effects. It represents a position of principle taken at a moment when Anthropic still dominated the enterprise segment, not a direct reaction to losing market share.

Third: regardless of the original intent, a company that builds its public identity around responsibility and caution cannot avoid the stricter standard it will inevitably be held to. When the call to slow down comes from one of the field's own technology leaders, and that same company is considering, just days later, an accelerated launch explicitly motivated by competition, the public has every right to ask where sincere caution ends and marketing begins — all the more so given that the stakes involve a tens-of-billions-of-dollars IPO, and "safety" has, in recent years, also become a selling point for corporate and institutional clients.

What this episode really says about today's AI market

Competition among leading AI companies has clearly intensified — market shares can shift by tens of percentage points within a few weeks, and a single successful launch, like GPT-6 Astra, can upend a balance held for years. At this pace, any voluntary pause — sincere or not — becomes costly almost instantly. Investor pressure, especially around major financial events like a stock listing, weighs just as heavily as arguments made in favor of safety.

Perhaps the discourse around safety isn't just marketing, but it doesn't guarantee it will be followed in practice either — it's more a form of strategic positioning, through which each company wants rules that benefit it in the long run, without giving up the advantage it holds right now. Anthropic seems caught between a rhetoric it chose to adopt publicly and a market that waits for no one — not even those calling, sincerely or not, for things to slow down.

While the leaders of major AI companies call for greater caution in developing frontier models, competition is pushing the technological race in the opposite direction.