Study Finds Anthropic's Internal AI Applications Double R&D Speed
nashnova research
Independent research group METR reports that Anthropic's R&D pace has roughly doubled since the company began using AI tools internally at scale, with about 80% of the acceleration attributed to AI — but the same report warns that a self-improving AI loop could produce "extreme and dangerous capabilities."
R&D speed doubled — how did they measure that?
METR interviewed roughly 15 Anthropic employees and obtained model training data, benchmark scores for unreleased models, and internal analyses.
The finding: as of October 2025, Anthropic's model-improvement rate was about twice the pace seen from March 2024 to October 2025.
This means → the number did not come from Anthropic's own PR. An outside group got inside data and ran its own estimate — that makes it a notch more credible than a company self-report.
80% attributed to AI — what exactly is AI doing?
Of the doubling, about 80% traces directly to AI use inside the company, not to hiring more people or adding hardware.
In plain terms = Anthropic put its own AI to work on its own R&D pipeline. The result: AI is helping humans build stronger AI.
METR adds, however, that AI research at Anthropic cannot yet run fully on autopilot. There are only "signs of life" suggesting further acceleration is technically feasible.
Who is METR, and how deep was its access?
METR is an independent research group focused on AI safety evaluation. This report was separate from the embedded-evaluation partnership that Anthropic CEO Dario Amodei had previously committed to.
Yet Anthropic granted METR internal-information access that "exceeded expectations" — including three unreleased models trained since July 2026.
This means → Anthropic voluntarily opened more data to an outside body than anyone expected. That is itself a signal: it wants third-party validation of its progress.
The self-improvement loop — where is the risk?
METR warns that if AI keeps improving itself in a feedback loop, the cycle could eventually produce "extreme and dangerous capabilities."
Put simply = today AI helps humans do R&D, which builds a stronger AI, which helps build an even stronger AI — once that loop starts running, it could accelerate faster than humans can erect safety guardrails.
This reflects a core tension in AI safety: accelerating R&D is itself the single largest source of risk — and Anthropic sits at the dead center of that contradiction.
市场有风险，内容仅供研究参考，不构成投资建议。