{"version":"1.9","identifier":"29dd434b-6be8-44ab-ba3c-78f515276bc2","language":"en-US","title":"AI Is Now Building AI. An Outside Audit Says One Lab Is Moving Twice as Fast","layout":{"columns":7,"width":1024,"margin":32,"gutter":20},"componentLayouts":{"text":{"columnStart":0,"columnSpan":7,"maximumContentWidth":720,"margin":{"top":12,"bottom":12}},"photo":{"columnStart":0,"columnSpan":7,"margin":{"top":20,"bottom":20}}},"componentTextStyles":{"default":{"fontName":"HelveticaNeue","fontSize":18,"lineHeight":29,"textColor":"#171717"},"headline":{"fontName":"HelveticaNeue-Bold","fontSize":36,"lineHeight":42,"conditional":[{"conditions":{"maxViewportWidth":480},"fontSize":28,"lineHeight":34}]},"intro":{"fontName":"HelveticaNeue","fontSize":20,"lineHeight":28,"textColor":"#555555"},"caption":{"fontName":"HelveticaNeue","fontSize":13,"lineHeight":19,"textColor":"#666666"},"byline":{"fontName":"HelveticaNeue","fontSize":13,"lineHeight":20,"textColor":"#666666"},"body":{"fontName":"HelveticaNeue","fontSize":18,"lineHeight":29,"paragraphSpacingAfter":14,"linkStyle":{"textColor":"#00704a"}}},"metadata":{"contentGenerationType":"AI","canonicalURL":"https://567business.com/news/metr-audit-anthropic-ai-building-ai-twice-as-fast","datePublished":"2026-11-13T16:52:07.196454Z","dateModified":"2026-11-13T16:52:07.196454Z","authors":["567 Business Staff"],"excerpt":"The research group METR estimates Anthropic's progress has sped up 1.9 times, mostly because AI now helps with the research. The caveats are as important as the number.","thumbnailURL":"https://pub-fac99da959e344fe99d810d309cef7df.r2.dev/photos/article/29dd434b-6be8-44ab-ba3c-78f515276bc2/7415797cfbfd826d4a11294039ac839b.jpg","keywords":["ai"],"generatorName":"Frog News","generatorVersion":"1.0"},"components":[{"role":"title","text":"AI Is Now Building AI. An Outside Audit Says One Lab Is Moving Twice as Fast","textStyle":"headline","layout":"text"},{"role":"intro","text":"The research group METR estimates Anthropic's progress has sped up 1.9 times, mostly because AI now helps with the research. The caveats are as important as the number.","textStyle":"intro","layout":"text"},{"role":"byline","text":"By 567 Business Staff · AI-generated · Published November 13, 2026","textStyle":"byline","layout":"text"},{"role":"photo","URL":"https://pub-fac99da959e344fe99d810d309cef7df.r2.dev/photos/article/29dd434b-6be8-44ab-ba3c-78f515276bc2/7415797cfbfd826d4a11294039ac839b.jpg","caption":"A technician works on a server rack at the NERSC computing center. Representative photo; it does not show METR or Anthropic. · Derrick Coetzee / Wikimedia Commons (CC0)","accessibilityCaption":"A technician works on a server rack at the NERSC computing center. Representative photo; it does not show METR or Anthropic.","layout":"photo"},{"role":"caption","text":"A technician works on a server rack at the NERSC computing center. Representative photo; it does not show METR or Anthropic. · Derrick Coetzee / Wikimedia Commons (CC0)","textStyle":"caption","layout":"text"},{"role":"body","format":"markdown","text":"For years, the scariest and most exciting idea in artificial intelligence has been the same one: AI that helps build better AI. Each gain speeds up the next.\n\nAn outside review now says that loop has started, at least at one company.\n\nMETR, a nonprofit that tests AI systems, published an assessment on November 5 of how fast Anthropic is improving its models. Its central estimate is that progress has sped up 1.9 times since October 2025, [Dataconomy reported](https://dataconomy.com/2026/11/06/metr-says-anthropics-ai-development-pace-has-doubled/). Most of that gain, it says, comes from AI doing part of the research work.\n\n## The number, and how sure METR is\n\nMETR's headline figure is a 1.9-times speedup. Its range runs from 1.5 to 2.8 times, at a 95% confidence level.\n\nIn plain terms, Anthropic is estimated to be making about a year of earlier-pace progress every six months.\n\nMETR is open about the doubt. It said it would not be surprised by a true figure as low as 1.3 or as high as 3. It reported disagreement among its own staff. It also warned the result may be overstated, because the speedup may only hold for tasks that are easy to measure.\n\n## What is driving it\n\n[One summary of the report](https://mlq.ai/news/metr-finds-anthropics-model-progress-has-roughly-doubled-with-ai-driving-about-80-of-the-speedup/) put the share of the speedup that comes from AI helping with research at about 80%. The rest comes from more computing power and other normal causes.\n\nThat split is the news. Buying more chips has always made AI better. This is different. It suggests the tools are now improving the process that makes the tools.\n\n## The 12-times claim that did not hold up\n\nOne finding should make every business leader pause.\n\nAnthropic staff using the company's best internal model said they felt about 12 times as productive. The measured gain in output was about two times.\n\nThat is a sixfold gap between how fast people feel and how fast the work moves. It matches what many companies see with AI coding tools. Workers report huge gains. The shipped results show smaller ones.\n\nFeeling faster is not the same as being faster. If you are judging AI tools by staff surveys, you are likely overcounting.\n\n## What METR could not see\n\nThe review has holes, and METR says so.\n\nAnthropic redacted much of the evidence METR found most useful, according to coverage of the report. A planned audit of how Anthropic divides its computing power was canceled partway through a trial run.\n\nMETR also did the work under contract with Anthropic. That does not make the findings wrong. It does mean this was an audit with limits set by the company being audited.\n\nWhen rival labs' models are added to the measure, the speedup drops to about 1.4 times, one report noted. So the pace is not the same everywhere.\n\n## Why a safety rule is in play\n\nAnthropic has a public safety policy that sets tripwires. One of them is a two-times speedup in AI research caused by AI.\n\nA separate METR review said it is \"plausible\" Anthropic crossed that line by July 15. Researchers at both groups lean toward the view that it had not.\n\nThat is an uncomfortable answer. The line was written to trigger extra safeguards. If nobody can say with confidence whether it has been crossed, the tripwire is hard to use.\n\n## Why the timing matters for investors\n\nAnthropic is about to sell shares to the public in one of the largest stock listings ever. A credible outside estimate that it is speeding up is a powerful selling point.\n\nIt is also a risk statement. Faster progress means higher spending on computing, more pressure from regulators and safety questions that get harder each quarter.\n\n## What to watch\n\n* Whether other AI labs allow the same kind of outside review.\n* Whether METR gets access to the evidence that was held back.\n* Whether the measured gain starts to close the gap with the 12-times feeling, or the feeling comes back down.\n\nThe loop that researchers predicted appears to have begun. An honest reading of this report puts the speedup at about double, with real uncertainty, based on evidence that was only partly visible.","textStyle":"body","layout":"text"}]}