Anthropic says AI is now speeding up its own research, a safety alarm it reported in its IPO filing

Anthropic says it has crossed one of its own safety lines. In an amended prospectus filed Tuesday, the company told would-be investors: "We made a determination that we have met this threshold." The threshold is the point where AI is doing enough of Anthropic's own research to roughly double how fast its models improve. The same filing set the price range for its IPO (initial public offering, the first sale of shares on the stock market) at $70 to $78 per share. That values the company at up to about $1.4 trillion, counting every share that could exist, according to Yahoo Finance.

The threshold sits in Anthropic's Responsible Scaling Policy, the company's rulebook for adding tougher safeguards as its models grow more capable. By the policy's own definition, crossing it means packing two years of AI progress, at the 2018-2024 pace, into a single year. The filing says AI "has already been accelerating our research and development." One detail matters a lot. Anthropic says it based the call on "what we expect from future models and our overall rate of progress, rather than on what we have observed from current models." Only a week earlier, it had told investors the trigger might come "imminently." Crossing the line means more safeguards, not a shutdown.

Independent evidence points the same way, and suggests it happened sooner. METR, a nonprofit that tests AI systems, estimated last week that Anthropic's models have been improving nearly twice as fast since October 2025. In a separate review, METR called Anthropic's August argument that it was still below the threshold "overstated." It said it was "plausible" the line had been crossed by mid-July. The underlying fear is old and well known. AI could improve itself faster than humans can check it. In 2023, CEO Dario Amodei joined other AI leaders in saying the "risk of extinction from AI" should be a global priority alongside pandemics and nuclear war. Anthropic now says it does not expect to safely use Claude "to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding." It adds that it is weighing policy changes "to keep humans in control."

Our thesis is that this is the first time an AI safety threshold has been reported to the market as a material fact, and that is a good thing. We have argued that the S-1, the registration document a company files before listing shares, is becoming the most detailed public record of a frontier lab. This filing proves the point. A commitment that used to live in a voluntary policy now sits in a document that securities regulators scrutinize and that investors can sue over. That gives it far more weight than a blog post. Anthropic deserves credit for declaring the threshold based on its forecast of future models, not only on what current ones do. That is the right standard for a risk that grows exponentially.

Still, two caveats. First, the timing. If METR is right, the line was plausibly crossed around mid-July, and Anthropic was still arguing in August that it had not been. Self-assessment lagged outside measurement by weeks. That fits a lesson we keep repeating: lab self-evaluation is not enough, and independent auditors like METR are essential infrastructure, not a nice extra. Second, the remedy is still vague. "Weighing policy changes" is not the same as specifying them. Investors and the public should ask which safeguards are now in force, who verifies them, and what would trigger a slowdown.

Our reading is that faster research is the very engine of our long-term optimism. AI that speeds up science is how we get to curing diseases and to material abundance. But that engine now runs inside a company about to be owned partly by retail savers, at a valuation of up to $1.4 trillion. Commercial pressure and caution will pull in opposite directions in every quarterly report. The near term will test whether market discipline reinforces safety governance or erodes it. The deciding question is not whether AI accelerates AI; per Anthropic itself, it already does. It is whether human oversight can scale at the same pace. Watch what the final prospectus and the next policy update actually commit to.