Metr's R&D Automation Index: How Much AI Already Builds AI Anthropic's Claude now leads 26% of Anthropic's internal model R&D tasks, up from under 1% in February, according to Metr's R&D automation index. Metr found Anthropic's internal capability scores improving at roughly 1.9x the earlier rate since October of last year (plausible range 1.5x to 2.8x), with about 80% of that acceleration attributed to AI itself, and Anthropic stated it has crossed its own "two times acceleration" safety threshold in its responsible scaling policy. Anthropic's report "When AI Builds Itself," co-authored by Jack Clark and Marina Favaro, says more than 80% of code merged into its codebase was written by Claude, while the underlying evidence is partly redacted, leaving independent verification dependent on Metr's and Anthropic's own characterizations. Metr's R&D Automation Index: How Much AI Already Builds AI Metr's data shows Anthropic's AI now leads 26% of its own research tasks. Here's what the numbers actually say about AI building AI. What is Metr’s R&D Automation Index? Metr’s R&D automation index is a scoring system that tracks how much of Anthropic’s AI model development work is being done by AI itself rather than by human researchers. Tasks are rated on a scale from no AI involvement, to AI assisting, to AI collaborating, to AI leading, up to fully automated. As of the latest reporting, Claude leads 26% of Anthropic’s internal model R&D work, up from under 1% in February. That six-month jump is the single number driving a wave of revised predictions about how close AI is to automating its own research. TL;DR - Anthropic’s AI now leads about a quarter of its own R&D tasks , according to Metr’s automation index, a jump from under 1% roughly six months earlier. - Metr found Anthropic’s internal capability scores are improving at roughly 1.9x the earlier rate since October of last year, with a plausible range of 1.5x to 2.8x, and Metr estimates around 80% of that acceleration comes from AI itself. - Anthropic says more than 80% of code merged into its codebase was written by Claude , and in one safety research task, Claude agents recovered 97% of a performance gap that took two human researchers a week to partially close. - Anthropic formally stated it has crossed its own “two times acceleration” safety threshold , a tripwire in its responsible scaling policy meant to trigger stronger safeguards when AI meaningfully speeds up its own development. - Task time-horizon length, how long a task an AI agent can complete unassisted, is now doubling roughly every 1.7 months by some estimates, compared to roughly every 7 months from 2019 to 2025. - Several prominent AI forecasters, including podcast host Dwarkesh Patel and the authors of the AI 2027 scenario, revised their timelines shorter after reviewing Metr’s report, with some now expecting a human-level AI researcher as early as 2027. - The underlying evidence is partly redacted , so independent verification of the exact numbers depends on trusting Metr’s and Anthropic’s own characterizations. › Build a safety-inspection app for 60 plants. already included: ✓ SOC 2✓ role-based access✓ audit log✓ self-hosted What exactly did Metr measure? Metr is an independent organization that evaluates frontier AI capabilities, often working directly with lab data under controlled access. For this report, Metr examined Anthropic’s internal records on how fast its model capabilities have been improving and compared the pace before and after roughly October of last year. The headline number is a 1.9x acceleration in Anthropic’s internal capability growth rate, with Metr giving a plausible range between 1.5x and 2.8x given the uncertainty in the data. Metr also estimated that a large share of that increase, around 80%, is attributable to AI systems themselves contributing to the research process rather than to other factors like more compute or more hires. That acceleration estimate is separate from, but related to, the R&D automation index. The index tracks what fraction of discrete R&D tasks writing code, running experiments, picking the next research direction are being done with AI leading or fully automating the work, rather than how much faster things are moving overall. How much of Anthropic’s research is actually automated? The automation index put Claude at leading 26% of Anthropic’s model R&D work in the most recent measurement, compared to under 1% in February. Anthropic’s own report, “When AI Builds Itself,” co-authored by Jack Clark and Marina Favaro, adds supporting detail: more than 80% of code merged into Anthropic’s codebase was reportedly written by Claude. The report also describes a specific test on an AI safety research problem. Two human researchers working for about a week recovered roughly 23% of a known performance gap. Claude agents working on the same problem recovered about 97%. In a separate test of research judgment, where the question was whether Claude could pick a better next research step than a human would, Claude Opus 4.5 won that comparison 51% of the time, and a newer preview model won 64% of the time. Those numbers matter because they move beyond “AI writes code faster” into “AI makes better calls about what to try next,” which is a harder and more consequential kind of automation. What is the responsible scaling policy threshold, and why does crossing it matter? Anthropic’s responsible scaling policy includes a defined tripwire for AI-driven research acceleration: it triggers when AI is speeding up Anthropic’s own progress by around two times, substantially because of automation rather than other factors. In a November amendment to its IPO filing with the SEC, Anthropic stated it has met that threshold. The filing’s language is careful. Anthropic says the determination is based on its internal estimates and its expectations for future models and its overall rate of progress, not solely on a backward-looking measurement of a model that has already crossed the line. The filing also includes a notable admission: Anthropic says its evidence about AI-driven acceleration comes mostly from lagging indicators, meaning it may not recognize a significant acceleration until after it has already occurred. Build an onboarding app for 4,000 new hires. Build ↵ Self-hosted SOC 2 Role-based access Audit log Days after the filing, Anthropic’s founders published a letter ahead of the IPO stating that building AI safely may require the company, and ideally the rest of the industry, to deliberately moderate the pace of AI capability growth, especially when that growth is driven by AI systems improving themselves. They wrote that there may be times when they choose to advance more slowly than they could, even at a competitive disadvantage. Is the acceleration showing up anywhere else? Yes. OpenAI published findings in September about internal usage of coding agents by its own researchers. The heaviest-usage 10% of OpenAI researchers went from spending about $1,000 a day on internal coding agents at the start of summer to over $7,000 a day by the end of August, a sevenfold increase in about three months. That kind of spending increase is a proxy for how much of the research workflow is being handed to agents rather than done manually. Separately, Metr’s long-running measurement of AI “time horizons” how long a task an agent can complete autonomously before failing has shown accelerating doubling times. From 2019 to 2025 that measure doubled roughly every seven months. From 2024 into early 2026 it doubled roughly every 3.3 months. Estimates following the release of Claude Opus 4.5 put the doubling time at roughly 1.7 months. Metr’s last officially measured model, an early Claude preview, scored around 3.1 hours on this metric. An unreleased internal Anthropic model, referred to in some analyses as “model 4,” is estimated at around 38 hours, or close to a full work week of unaided task completion. Metr has reportedly stopped updating its public time-horizon chart because current models have outpaced what the existing test can measure. Why did so many AI skeptics change their minds? The report landed hardest because of who reacted to it. Podcast host Dwarkesh Patel, who had argued a year earlier that AGI wasn’t imminent because models couldn’t learn on the job, published a widely viewed essay reversing much of that position, saying a human-level AI researcher could arrive within a couple of years, possibly months. The AI Futures Project, the team behind the AI 2027 forecasting scenario, published an update stating that real-world progress now looks roughly five months ahead of their original scenario’s timeline. Ryan Greenblatt, a researcher affiliated with Metr, wrote that he had previously expected only a 1.15x to 1.3x acceleration in 2026 and called the actual 1.9x figure, and his own rough extrapolation of roughly 4.5x acceleration specifically in AI software progress, a “big update” toward the possibility of superintelligence within one to two years. The common thread in all these reactions is Dwarkesh’s broader argument: AI research, like mathematics and competitive programming before it, is unusually verifiable. You can measure whether a training trick lowered loss, sped up inference, or cut memory use. Verifiable domains are exactly where AI systems have shown the fastest jumps, going from mediocre to superhuman in a few years for math competitions and coding contests. If AI research shares that property, progress in automating it could compound quickly rather than gradually. How reliable are these numbers? Enterprises build with AI on Remy. - Self-hosted - SOC 2 - HIPAA - Audit log - Role-based access Less certain than the headlines suggest. Metr’s report relies heavily on data provided by Anthropic, and much of the underlying evidence was redacted before publication, meaning outside observers are trusting Metr’s interpretation rather than verifying raw numbers themselves. Greenblatt, despite finding the report convincing, noted explicitly that less redaction would reduce how much trust is required. Extrapolations built on top of the automation index, such as projecting when 95% of R&D tasks might be fully automated, are speculative trend-fitting exercises, not measured outcomes. As of the most recent reporting, no task had actually been rated “fully automated” on Anthropic’s five-level scale. Frequently Asked Questions What does “AI leads 26% of R&D” actually mean? It means that, under Anthropic’s own task-rating system, roughly a quarter of discrete model development tasks are currently categorized as AI-led rather than human-led or AI-assisted. It does not mean 26% of all research output or company activity is autonomous. What triggered Anthropic’s safety threshold under its responsible scaling policy? The threshold triggers when AI is accelerating Anthropic’s own research and development by roughly two times, and that speedup is substantially caused by automation. Anthropic stated in a November SEC filing amendment that it has met this threshold, based partly on expectations for future models rather than solely retrospective measurement. Is recursive self-improvement RSI already happening? Partially, by some definitions. AI systems are measurably speeding up AI research and contributing a large share of code and research decisions. But full, autonomous recursive self-improvement, where AI drives its own progress with no human bottleneck, has not been reached; Anthropic’s own five-level scale shows no task yet rated “fully automated.” Why do verifiable domains like math matter for this story? AI systems improve fastest in domains where answers can be automatically checked, since that allows massive repeated practice. Math and competitive programming went from weak to superhuman performance within a few years once this dynamic kicked in. The argument driving recent timeline revisions is that AI research itself has similar verifiable properties did the loss go down, did the model get faster , which could let automation compound in a similar way. Can these numbers be independently verified? Not fully. Much of the supporting data in Metr’s report was redacted before release, so the public is relying on Metr’s and Anthropic’s own summaries and characterizations rather than raw figures. Researchers who found the report convincing, including those at Metr, have acknowledged this limitation directly.