Anthropic · AI R&D · Recursive improvement
Claude now leads a quarter of the work building Claude's successors
Anthropic says Claude leads 26% of the AI R&D work inside the company and collaborates on more than 90%. Its “lead” level means the system can take a high-level task and complete most of it end-to-end while a human supervises; it does not mean Claude independently designs, trains and releases a quarter of a new model.
The company says the lead share was under 1% in February. It also says none of the measured R&D work reaches full autonomy. That distinction is the whole story: this is neither a vague anecdote nor a runaway loop. It is a company-produced measurement of AI taking a meaningful share of frontier AI research under supervision.
Anthropic reports roughly 30,000 research-and-engineering agents active at any one time on its most-used internal platform in August. Every action passes through an online monitor before execution; the company says about one in 47,000 decisions was blocked. The methodology still deserves outside scrutiny, especially because Anthropic uses AI systems to help evaluate AI work. The direction, however, is difficult to dismiss.
Source: Anthropic ↗Anthropic · Biology · Physical automation
Claude has left the simulation and entered the wet lab
Anthropic has quietly set up a physical biology laboratory in the San Francisco Bay Area. Its head of life sciences confirmed the lab to Reuters and said the company believes biology's final test remains real laboratory work, not computer simulation alone.
Reuters reports that Anthropic wants to explore Claude directing robotic units that carry out science experiments with limited human intervention. The company says human oversight and involvement remain essential for safety, and describes the work as early. Still, this is a distinct step: a coding agent acts on files and networks; a laboratory agent can eventually act on matter.
Anthropic says it is focused on hard-to-treat or economically neglected conditions and is not running clinical trials for now. The dual-use tension remains obvious. More capable scientific systems may speed valuable research, while expanding the need to control what powerful systems are asked to investigate.
Source: Reuters ↗Mantic · Forecasting · Prediction
The AI entered a forecasting contest and humans discovered they were the baseline
London startup Mantic announced a $25 million seed round after its AI forecasting system outperformed human participants in the summer 2026 Metaculus Cup. Reuters reports that AI entrants collectively dominated the competition, which asks forecasters to assign probabilities to future political, economic and cultural events.
Mantic does not claim to train a frontier model from scratch. It specializes existing frontier models for forecasting, tests them against historical questions, grades the forecasts and iterates. Forecasting is a useful stress test because it rewards calibrated uncertainty, not simply a fluent answer.
One tournament does not create an oracle. These systems can inherit weak assumptions from their models and their data, and success on defined questions does not mean they predict arbitrary world events. But even a small, well-calibrated advantage has obvious value in finance, policy and planning—hence the seed round arriving right on schedule.
Source: Reuters ↗