METR: Anthropic's AI is speeding up its own research about 2x — and most of the gain is AI itself
METR, an independent AI safety research organisation, published an assessment on November 5 after 11 weeks inside Anthropic: the company's models are improving about 1.9 times as fast as they were before October 2025 — and roughly 80% of the acceleration comes from AI use itself, not from more computing power. Anthropic is private, so Google and Amazon — both Anthropic investors — are the public read-throughs. One caveat: Anthropic redacted most of the evidence METR found most informative.
📊 Scorecard: our Bullish call on $GOOGL is a miss so far: from $288.15 to $287.38 at the latest close (-0.3%). Against QQQ the call is right: the stock is 0.4 points ahead of the index.
The News
METR, an independent AI safety research organisation, published on November 5 an assessment of AI-assisted research at Anthropic after 11 weeks of privileged access — August 12 to October 29, with about two and a half full-time researchers, internal benchmark data, training-compute metrics and around 15 staff interviews. Its headline finding: benchmark scores for Anthropic's models, measured on Anthropic's own capability index, have been growing at roughly 1.9 times the pre-October 2025 rate, with a 95% confidence interval of 1.5 to 2.8 times. METR describes that as roughly two years of historical progress over the last year. A confidence interval is the range the true figure probably falls in — here, the acceleration is almost certainly real, even if its exact size is uncertain.
METR says it is confident the jump did not come from scaling training compute — buying more computing power to train bigger models — and estimates, with substantial uncertainty, that about 80% of October's acceleration traces to AI use itself, a share that has grown over time. In plain terms: the models are now doing a large share of the research work that makes the next models better. Three of Anthropic's newest internal models, labelled Model 2, 3 and 4, finished training between July and October 2026 and are already contributing to the gains; they have not shipped publicly. METR saw no sign of bottlenecks in compute, data or algorithmic progress over the next 12 months.
The caveat is stated plainly. Anthropic redacted a majority of the evidence METR found most informative, so outside readers cannot check the mechanics — only METR's summary judgment. Internal Slack polls cited in the report found Anthropic staff self-reporting being about 12 times as productive with the company's most capable internal model, roughly double what they reported a few months earlier with the internal model Mythos 5. That figure is self-reported, not measured output. Anthropic is a private company — for now. It filed a public S-1 on October 26, amended November 2, to list on Nasdaq as ANT, with the roadshow able to start November 10 and a listing targeted for the week of November 16. Until then, Google and Amazon, both Anthropic investors, are the public read-through tickers for this story.
Anthropic's AI is now doing most of the work of making Anthropic's AI better — but the evidence behind that claim is largely redacted.
Why It Matters
This matters because it changes the shape of the AI race. Until now, progress in AI has mostly come from spending more — more computing power, more data, more engineers. If a lab's models are now doing the research that makes the next models better, the returns on having the best model compound: the leader's advantage grows not just because it has more money, but because its own AI is working on the next version while its rivals' teams are still doing that work by hand. That is the feedback loop the industry has talked about for years, and METR's assessment is the first inside look suggesting it is actually happening.
It matters for investors, too, because the read-through is broad. Google and Amazon both back Anthropic, and both sell the cloud computing that AI labs buy. Faster model progress means more capable models sooner, more products built on them, and more computing sold to run them. The 12-times self-reported productivity figure, if even half true in measured output, would mean a research team of ten now does the work of a much larger one — which is why the labs are hiring less and deploying more of their own models instead.
And it matters with its limits stated. The redaction is not a footnote — it is the reason this story is a judgment, not a proof. Anthropic chose what METR could show, and what it hid was, by METR's own account, the most informative evidence. The numbers are the best outside look anyone has had inside a frontier lab, but they are still one organisation's summary of evidence it cannot fully share.
Analyst's Take
Our call is Bullish. If METR's numbers hold, the AI industry has crossed a threshold that matters more than any single product launch: the research loop is now partly self-driving. A 1.9-times acceleration in capability growth, with roughly 80% of it coming from AI use rather than more computing power, means the constraint on progress is shifting from hardware budgets to how well a lab can deploy its own models on its own problems. That is good news for the labs with the best internal models — and for the cloud and chip companies that sell to them, since faster progress means more demand, not less.
What to watch: whether other labs report anything similar, and whether Anthropic ever un-redacts the evidence. The redaction is the honest limit on this story — METR had the access and the rest of us have METR's word. The 12-times self-reported productivity figure is a survey answer, not a measurement. But the direction of travel is corroborated: The Verge's review of the findings reports the same roughly twofold acceleration, driven mainly by AI use rather than more computing power. If the self-improvement loop is real, the next twelve months should show it in shipped models — and METR says it saw no bottleneck that would stop that happening.
Are you bullish or bearish on this story?
By The Numbers
FAQ
What is METR, and why does its word carry weight?
METR is an independent AI safety research organisation. It spent 11 weeks inside Anthropic — August 12 to October 29 — with privileged access to internal benchmark data, training-compute figures and about 15 staff interviews. That inside access is what makes the report unusual; most outside assessments can only guess at what happens inside a lab.
What does 1.9 times faster actually mean?
METR measured Anthropic's models on the company's own capability index and found benchmark scores growing at roughly 1.9 times the pre-October 2025 rate — METR describes it as roughly two years of historical progress over the last year. The 95% confidence interval is 1.5 to 2.8 times, which means the true figure is probably in that range.
What does it mean that 80% of the gain is AI itself?
METR estimates — with substantial uncertainty — that about 80% of October's acceleration traces to AI use itself, a share that has grown over time, rather than to buying more computing power. In plain terms: the models are now doing a large share of the research work that makes the next models better, instead of humans or bigger computers doing it.
What is the redaction caveat?
Anthropic redacted a majority of the evidence METR found most informative. Outside readers cannot check the mechanics — only METR's summary judgment. The headline numbers are METR's, but the workings behind them are hidden.
What is the 12x productivity figure?
Internal Slack polls cited in the report found Anthropic staff self-reporting being about 12 times as productive with the company's most capable internal model — roughly double what they reported a few months earlier with the internal model Mythos 5. It is self-reported, not measured output, and it should be read that way.
👁
Source: METR via Pondero