Patel Says Anthropic Contact Forecast AI Could Beat Human Researchers in Nine Months

Dwarkesh Patel said an unnamed Anthropic contact expects AI systems to surpass human researchers in AI research within nine months, a forecast that the company’s public index does not directly measure.

Patel recounted the conversation in a podcast episode published Oct. 15 at 5 a.m. ET (09:00 UTC). He quoted the contact saying, “I expect A.I.s to beat humans at A.I. research within nine months.” Patel said the person cited inside information but declined to identify a model, evaluation or other evidence.

Anthropic’s index tracks the role of its Claude models across the company’s AI research and development tasks, including the level of human oversight. A Sept. 17 report presented an August 2026 snapshot in which Claude led 26% of the work measured. AI was involved at level 3 or above in more than 90% of those tasks, while the index recorded no work at level 5.

At level 3, AI works on substantial parts of a task with close human direction. At level 4, an AI system can complete most of a task from a high-level prompt while a person supervises. Level 5 represents fully autonomous work. These categories describe the work process and degree of oversight, not an AI system’s performance against human researchers.

That distinction leaves the nine-month forecast separate from the index’s findings. Patel did not name the person or explain what they meant by beating human researchers, and the public metrics do not identify a benchmark or test that would resolve the claim.

The index is a prototype and uses Anthropic models to assess Anthropic systems. The company also notes that AI labs do not share a common method for comparing these measures, limiting comparisons across companies. The August results describe Anthropic’s own R&D workflow rather than a broad measure of AI research ability. The stated purpose is to track AI’s role in developing future systems and whether people can oversee that work as automation increases.

Anthropic has warned that further automation of AI development could make systems harder for people to understand or control, a risk that sits alongside the potential for faster model development.