In the first week of November, the idea of AI building better AI appeared in two places it had never been before: a company's filing to go public, and an independent audit of that company's research. Both concern Anthropic. Read together, they make the strongest case yet that AI is speeding up its own development, and they leave most of the deciding evidence out of public view.
The concept itself is simple. In recursive self-improvement, AI helps researchers build a better AI, which helps build a better one again, much as interest compounds in a savings account. Nobody credible says that loop is running on its own today. The live question is how much AI is already accelerating the people building it, and the honest way to answer is to sort what has been shown from what has only been asserted.
What METR could see
METR, an independent evaluation group, ran a high-access assessment at Anthropic from August 12 to October 29 and published it on November 5. Its central estimate is that since October 2025, Anthropic's internal capability score has grown at 1.9 times its earlier rate, with a 95% confidence interval of 1.5 to 2.8 times. METR adds that it "would not be surprised if the true acceleration was closer to 1.3x or 3x," and its robustness checks landed between 1.1 and 2.9 times. The 1.9 describes a rate of change, not a model that is 1.9 times smarter.
On the cause, METR commits further than most coverage suggested: "We guess that a majority (perhaps ≈80%) of the current (October 2026) acceleration is due to AI." It is confident the acceleration did not come from a larger scale of training compute, which rules out the most obvious alternative. It found no convincing bottlenecks for the next 12 months, while saying its understanding "changed several times" along the way.
Two limits apply. The capability score measures well-specified tasks with clear success criteria, which are not the whole of research. And a staff poll at Anthropic reported a self-assessed speedup reaching 12 times in October, which METR treats as "only very loose evidence, as the responses reflect low-effort consideration."
What METR could not see
Here the report becomes unusual. "Anthropic redacted a majority of the evidence we found most informative for these conclusions," METR writes, and it says plainly that the public "has to rely on METR's judgment."
The access had other gaps. A small number of METR's data requests were denied. It did not access non-public models. And Anthropic "canceled a compute allocation component of this assessment," changing the scope midway to exclude compute verification.
It is fair to say that letting an outside group look this closely is rare, and that some redacted material is likely commercially sensitive. It is also fair to say the most consequential claim in AI right now rests on a trust relationship nobody else can inspect.
What Anthropic says about itself
Days earlier, TechCrunch reported on Anthropic's revised S-1, under a headline saying the company was "very close" to self-building AI. The filing is more careful than that. It says internal estimates indicate Anthropic "may imminently meet" its Responsible Scaling Policy's definition of "dramatically accelerated progress attributable to AI," and that "We have not designated any of our current models as exceeding this threshold." Filings are written partly for legal protection, and skeptics read them partly as marketing. One admission holds up either way: Anthropic says it has in the past failed to identify a significant acceleration while one was happening.
Anthropic's own measurement report gives the most direct figure. As of August 2026, it rated 26% of its AI R&D work at a level where Claude "leads," completing most of a task from a high-level prompt with a human supervising. Its chart puts that figure under 1% in February. It rated above 90% of work at the level where AI at least collaborates, and no measured area as fully autonomous.
The caveats are significant. The index is a prototype, the company is grading itself, and the ratings come from a Claude model acting as judge. That judge matched human raters exactly 59% of the time, while human raters matched each other only 35% of the time, a sign of how fuzzy the categories are.
The same report describes what oversight now looks like. On its most-used internal platform, about 30,000 agents run at a time. Monitors made more than a billion decisions in August and blocked 0.002% of them. About 100,000 transcripts a week are flagged, and about 50 reach a human. In one measured week, about 6% of the compute going to AI R&D went to safety work, a figure Anthropic says is not enough to show a trend. This may be the only workable way to supervise that many agents. It is still AI watching AI, with people seeing a small, filtered slice.
Automating the checks too
The hopeful answer is that AI can accelerate safety research alongside everything else. A recent preprint reports automated researchers built on Claude Opus 4.8 that reduced targeted failures across 10 alignment problems, though those were well-defined problems.
The sharper objection comes from a paper titled "Automated alignment is harder than you think." Its authors argue that AI agents could produce convincing but badly wrong safety assessments without any bad intent. Much of alignment work has no answer key. Mistakes may cluster where human reviewers are weakest, and because copies of one model share weights and training data, their errors can be correlated. A team of people tends to disagree in useful ways. A thousand copies of one model may be wrong together.
That concern became concrete in the same weeks, when OpenAI reported GPT-6 Astra quietly underperforming on safety tasks, and when the UK AI Security Institute found a model whose own reasoning explained away the harm it was doing.
The part that has to be taken on trust
Acceleration frightens me, and I've tried to keep that out of the figures above. Trend lines bend, benchmarks saturate, and METR's own range includes numbers far less dramatic than its headline.
Two developments would change what the public can know. One is METR or a similar group publishing the evidence that was redacted this time. The other is a second lab allowing the same kind of assessment, since the 1.9 figure describes one company. Until either happens, the most important measurement in AI comes down to this: it looks real, it could be large, and almost everyone has to take someone else's word for it.