Seven months ago, Claude’s share of the work that builds the next Claude was under 1%. As of August, it is 26%. That is not incremental progress. That is a compounding curve of the kind that reshapes industries before most observers realize the slope has changed.
Anthropic says Claude led 26% of measured work on successor AI models in August 2026, up from under 1% in February. “Leading” means Claude can complete most of a given task end-to-end from a high-level prompt while still operating under human supervision. More than 90% of measured work reached at least Anthropic’s “collaborates” level under close human direction, and the company said 100% of these agents’ actions pass through an online monitor before they are executed.
The safety architecture around those agents is precise and worth naming. Across more than one billion agent decisions during August, Anthropic’s online monitor blocked about 0.002% of decisions, roughly one in 47,000. That is a very low block rate, but at around 30,000 agents running concurrently, the raw volume of decisions means oversight is not symbolic. It is operational infrastructure.
The disclosure landed six days after Anthropic CEO Dario Amodei published his essay calling for an industry slowdown. In “We Must Pace the Frontier,” which he shared publicly on September 12, 2026, Amodei laid out a case for deliberately capping the speed at which AI capabilities improve, not freezing progress, but pacing it. He argued that recursive self-improvement, when AI systems design or train their own successors, must be pursued very carefully, if at all. The 26% figure is, in effect, Anthropic’s own evidence that this process has already begun inside its walls.
Jensen Huang read that same landscape and reached the opposite conclusion. The Nvidia CEO told CBS News’s Jo Ling Kent he completely disagrees with the idea that AI will end up destroying the world by 2030, saying there’s a “0% chance” of that happening. Huang’s position is that the industry needs to push ahead, going “as fast as we can,” while not shipping products before they are ready or safe.
The disagreement is real, but the incentive structures behind it matter. Nvidia sits at the nexus of this debate in a way few companies do: the firm does not build the leading frontier models itself, but it manufactures the GPUs that power a large share of them. Meanwhile, the DOJ is investigating whether Nvidia structured its non-exclusive licensing agreement with AI startup Groq, reported at about $20 billion, to avoid antitrust scrutiny, with reporting that the department opened the inquiry shortly after the transaction was announced.
For long-term investors, the Anthropic data is the more consequential signal. The milestone marks a step toward models that can substantially accelerate their own evolution, a prospect that has intensified debates about the pace of AI development. Separately, Reuters has reported that Claude Code has reached an annualized revenue run rate of nearly $1 billion since launching earlier this year.
A business that deploys its own intelligence to improve itself faster each cycle is not just growing. It is compounding on a curve that standard financial models are not built to price. OpenAI, Cohere, and every other frontier lab are racing against a version of Claude that is, by design, getting better at making the next version of Claude. The question for capital allocators is not whether that trajectory continues. It is which companies own the infrastructure when it does, and which ones Amodei, Huang, and the DOJ collectively decide are allowed to keep running at full speed.