By Jia-Nan Song, Tech Correspondent
Just a few weeks after claiming the title of the world’s most valuable AI startup, Anthropic dropped a massive open letter urging the entire planet to hit the brakes on frontier AI research.
On June 4, co-founder Jack Clark and head of research Marina Favaro penned a detailed report backed by hard internal benchmarks. They laid out exactly how fast “recursive self-improvement” (RSI) is actually taking off—and it’s moving way faster than anyone expected. That rapid acceleration is the very reason they’re pushing for a coordinated, global approach to managing where this tech is headed.
Let’s look at the numbers. By May 2026, Claude was autonomously generating 80% of the code that actually made it into production. Compared to mid-2024, our engineers were pumping out effective daily code at eight times the rate. Throw the latest Mythos Preview model into the mix, and research productivity jumped fourfold over baseline human-only workflows. We’re also seeing complex scientific tasks double in capability roughly every four months now—a massive drop from the seven-month cycles we used to track. In some brutal algorithm optimization scenarios, a human team might grind for hours to squeeze out a 4x speed boost. Mythos? It’ll crank out a 52x leap in one shot.
After tracking these metrics across multiple years, the conclusion is pretty clear: if this upward trajectory holds, and if AI systems start building out those spark-of-human creativity skills, we’re looking at a future where machines literally design and upgrade themselves. And it’s not just possible—it’s already knocking on the door.
“Picture this: progress won’t be limited by what we can build manually anymore. It’ll hinge entirely on compute availability or how quickly we crack new training and inference efficiencies. Our role shifts from hands-on builders to oversight—watching, testing, and auditing an ever-expanding ‘virtual lab’ run mostly by the AI itself. And once those automated R&D engines get their footing, they’ll spill over into other sciences and completely reshape them.”
That’s why Anthropic believes giving the world a legitimate option to slow down—or even temporarily halt—frontier AI isn’t just sensible, it’s necessary. It buys society and alignment researchers the breathing room they need to catch up with the tech. We’re rolling up our sleeves with partners to help draft the actual frameworks and guardrails that make a verifiable slowdown possible.
“These frameworks would let developers actually verify whether other labs worldwide are truly pulling back, while keeping bad actors from using ‘coordinated slowdown’ as cover to secretly race ahead.” The company stressed that if such a system goes live, Anthropic is ready to pull the brakes ourselves—but only if our closest competitors commit to the exact same verifiable pace.
All of this drops right in the middle of Anthropic’s biggest capital push yet. On May 28, the company closed a staggering $65 billion Series H round, shooting its post-money valuation past $965 billion and officially dethroning OpenAI’s $852 billion mark.
The backer roster reads like a dream team. Altimeter Capital, Dragoneer, Greenoaks, and Sequoia led the round, alongside $15 billion in pre-committed cloud infrastructure deals—including a $5 billion chunk straight from Amazon. Flash memory giants Micron, Samsung, and SK Hynix are also sitting in the investor chair.
In the official statement, Anthropic mapped out exactly where the money’s going: doubling down on AI safety and interpretability research, massively scaling compute infrastructure to keep up with Claude’s growing appetite, and expanding both the product suite and partner network.
Then there’s the IPO chatter. Back in March, reports surfaced that Anthropic was eyeing an October listing to raise north of $60 billion. More recently, insiders say the SEC quietly received confidential draft registration statements in early June, though share counts and pricing are still being hammered out. Morgan Stanley and Goldman Sachs are lined up as lead underwriters, with JPMorgan jumping in too, and word on the street is more banks could be added to the syndicate soon.
When asked about going public, co-founder and president Daniela Amodei kept it real: the sky-high cost of training next-gen models is leaving little choice but to tap public markets. “Building AI models is insanely capital-intensive,” she noted, adding that the public equity market is basically built for scenarios exactly like this.
Revenue is climbing fast, too. Earlier this year, Anthropic drastically revised its forward-looking projections, forecasting a fourfold sales jump to hit $18 billion this year, with a roadmap pointing toward $55 billion by 2027.
So here’s the kicker: on one side, you’ve got aggressive fundraising, a sprint to IPO, and heavy bets on commercializing models. On the other, a very public plea for the industry to cool its jets. That stark contrast has sparked a serious debate across the tech ecosystem.
Certain investors are calling it what it sounds like: classic regulatory capture. The theory goes that by constantly amplifying AI safety risks to push for stricter laws, Anthropic is essentially raising the compliance floor. That price of entry crushes smaller startups and open-source projects, letting Anthropic stretch its legs and lock in market dominance. Market analysts have echoed similar skepticism, pointing out the timing. Dropping a heavy risk assessment right before an IPO doesn’t just validate technical leadership to Wall Street—it also justifies a sky-high valuation while indirectly pressuring regulators into drafting rules that favor well-capitalized incumbents.
That said, plenty of academic voices think the alarm bells are ringing for good reasons. Dismissing safety warnings purely as a cash grab misses the point. Companies on the front lines of model iteration see cracks forming long before outside observers do. At the end of the day, we’re all trying to walk that tightrope between reaping the benefits of exponential tech and keeping runaway risk in check.
