The most striking sentence in the new report is not the headline number. It is the admission that sits right above it. Anthropic, the lab under study, held the right to remove private material from the public text, and the outside team says it used that right on most of the proof that mattered most.
That team is METR, a nonprofit group that tests frontier AI systems for risk. From August 12 to October 29, 2026, METR staff worked inside Anthropic on a simple question with large stakes. Is AI now speeding up the work of building better AI, and if so, by how much? Their report, Independent Assessment of AI R&D Acceleration at Anthropic in October 2026, was published on November 5.
The Bigger Story
Start with the number METR can show. The group tracked a score called AECI, short for Anthropic Epoch Capabilities Index. An index of this kind rolls many test scores into one number for each model, so a reader can see a trend across years of releases instead of cherry picking a single test.
Before October 2025, that index rose by about 15 points a year. After that point it rose by about 29 points a year. METR puts the change at 1.9 times the old rate, with a 95 percent confidence interval of 1.5 to 2.8 times. In plain terms, Anthropic made about two years of its old benchmark progress in a single year.
METR is careful to widen that picture. Its broader judgment is that progress runs at around twice the old rate, and it says it would not be surprised if the true figure sat closer to 1.3 times or to 3 times. Other ways of handling the same data gave results from 1.1 times to 2.9 times. The central number is an estimate with soft edges, and METR says so in its own text.
Then comes the cause. METR says it is confident the faster pace did not come from faster growth in training compute. Compute is the raw machine power used to train a model. If a lab simply buys far more of it, scores can rise without any change in how clever the research itself has become. METR looked at that path and ruled it out as the driver here.
What remains is the loop in the headline. METR guesses that a majority of the current acceleration, perhaps about 80 percent, comes from Anthropic using AI to help do its own research. That guess carries what METR calls substantial uncertainty. It rests on staff interviews, on surveys, on code and experiment data, and on evidence the public cannot see because Anthropic removed it.

Context
Recursive improvement has a precise meaning in this debate, and it helps to keep it narrow. The strong form would be a system that can design and build its own successor with no human in charge. METR does not claim that has happened. Its report says research at Anthropic is sped up mainly by AI assistance, not by fully autonomous work, though it notes some signs that autonomous contributions may play a larger role soon.
That gap matters. A researcher who writes code faster with AI help is not the same thing as a lab that runs itself. METR makes the point with its own survey data. In Slack polls inside Anthropic, staff put their personal speed gain from the best internal model at about 12 times in October 2026. METR calls that figure very loose evidence, and it warns in plain language that a personal speed gain does not equal a lab wide acceleration. One person moving faster does not make every stage of training, testing, and review move at the same pace.
The history behind the report also matters. Anthropic treats fast AI driven research as a risk signal in its own safety framework, because a lab that improves too fast may outrun its own checks. It invited METR in to test a claim from its August risk report, which held that its models had not yet crossed a line where two years of progress arrive in one. METR spent eleven weeks with internal test scores, training data, and staff interviews, working on Anthropic devices and taking no payment. A related safety argument, from a very different angle, runs through Lawsuit To Freeze AI Splits Safety Debate.
METR also found no clear bottleneck that would block further progress in the next twelve months. It checked algorithmic progress, training compute, experiment compute, and data, and found no convincing limit in any of them. It adds, with unusual candour, that its view on bottlenecks changed several times during the work, so that part of the report carries low confidence.
Anthropic redacted a majority of the evidence we found most informative for these conclusions.
Different Perspectives
One fair reading is alarm, and it is not hard to see why. If a better model speeds up research, and faster research yields a better model, the pace can compound. Each turn of the loop makes the next turn shorter. METR itself raises that risk, noting that a lack of clear bottlenecks means further progress could lead to even faster acceleration in the coming year. On this reading, the report is an early measurement of a process that safety work has warned about in theory for years.
A second reading is caution about the measure itself. AECI tracks tasks with clear checks, where a right answer can be scored. Much of real research is messier. Choosing which problem to work on, spotting a flawed result, or judging a strange failure mode does not score as neatly. METR accepts that its index may overstate or understate what matters, and that benchmark gains can reflect a lab getting better at the tests it tracks. The 1.9 figure describes growth on that index. It does not directly measure wisdom, taste, or judgment in the lab.
A third reading focuses on trust. Anthropic described its removals as aimed at removing private material that could speed up rival labs. That is a coherent business reason, and METR records it without calling it false. Still, the effect is plain. The public report asks readers to accept a large claim, about a large risk, on the word of a small outside team whose best proof stayed private. METR itself says the report does not achieve transparency about evidence, and that the public must rely on its judgment much more than would be ideal. Readers who doubt the claim and readers who accept it are both, in a sense, taking a position on trust rather than on data they can check.
There is also a quieter perspective from inside the work. Staff estimates in interviews centred around a doubling, but they ranged widely. Some of the gain in late 2025, METR says, likely came from causes that had nothing to do with AI at all. Hiring, data work, and plain fixes to old bottlenecks can all bend a trend line for a while. A trend that bends once for mixed reasons may bend back. Nothing in the report proves the new pace will hold, let alone keep rising.

Why It Matters
For people outside any lab, the first effect is timing. Safety reviews, laws, and public debate all assume some pace of change. If the pace inside a leading lab has doubled, calendars built on the old pace age fast. A review process that took a year to set up may now span two generations of models instead of one. That does not make the review worthless. It makes speed itself a policy fact, not just a technical one.
The second effect is about who gets to know. For years, debate about recursive improvement ran on forecasts and thought experiments, because no one outside a lab could see the work. This report changes the setting without settling the argument. An outside group has now seen the inside data and reported a doubling. Yet the same report shows the limit of that access. Seeing is not the same as showing. Until labs share more of the underlying evidence, public debate will lean on the judgment of a handful of groups with rare access, and everyone else will argue from summaries.
The third effect is competitive. If AI assisted research is now a real source of speed, then access to the best internal models becomes a research advantage, not just a product advantage. Labs with strong models can use them to build the next ones faster, while labs without them fall behind at a growing rate. Anthropic made that point itself, in its stated reason for the removals. Sharing exactly what drives the gain, it argued in effect, would hand rivals a map. That logic is hard to dispute and hard to love at the same time. It treats the method of acceleration as a trade secret at the very moment the acceleration becomes a public concern.
Readers should also keep scale in view. This is one lab, studied for eleven weeks, on one class of tasks. Other labs may be moving faster or slower, and METR offers no ranking. A doubling at Anthropic is a signal about the field, not a census of it. The signal is strong enough to take seriously and narrow enough to handle with care.
Closing Reflection
The strangest part of this story is the shape of the proof. A report about a machine helping to build its successor arrives with most of its best pages blacked out, and the black bars end up carrying as much meaning as the numbers.
That may be the honest picture of where things stand. Something inside at least one major lab appears to be moving about twice as fast as it did, largely because the product has joined the team that makes it. No one outside can fully check that claim, and no one inside seems fully sure where it leads. The loop, if it is a loop, has started quietly, in code reviews and experiments and late night runs, rather than in a single dramatic moment. The question the report leaves open is not whether AI helps build AI. It clearly does. The question is how fast that help compounds, and whether anyone watching from outside will be shown enough to know before the answer is obvious.
Reporting draws on the METR report Independent Assessment of AI R&D Acceleration at Anthropic in October 2026, published November 5, 2026, covering work from August 12 to October 29, 2026, checked November 16, 2026.
Sources
- METR, Independent Assessment of AI R&D Acceleration at Anthropic in October 2026, November 5, 2026: https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/
- Illustrations: Dirty Static
END OF TRANSMISSION
Stay curious. Question everything.


