Anthropic is using AI to build AI, doubling its research speed
METR's investigation finds that roughly 80% of Anthropic's current research acceleration comes from using its own AI to conduct research, which is precisely the early sign of recursive self-improvement that observers have warned about for years.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
Anthropic's research speed has doubled year-on-year
After 11 weeks examining Anthropic's internal test scores, training data, and employee interviews, METR concluded that as of October, Anthropic's speed on quantifiable tasks had roughly doubled from a year prior, with a confidence interval of 1.3 to 3x. One piece of evidence: plotting all of the company's model test scores (including three unpublished internal models) on a timeline, the slope suddenly doubled from October last year. METR also ruled out increased compute as the main driver—while compute has indeed increased, it cannot explain this acceleration.
Two years' progress compressed: what doubling really means
By late 2023, Claude was just a chatbot that couldn't read images or browse the web; by late 2025, it could work independently for over 30 hours and build a complete Slack-like application without human intervention. That's two years of progress. Then just one year later, AI helped solve a math problem that had puzzled mathematicians for 87 years. This is what METR means by walking through two years of progress in one year.
80% of the acceleration comes from AI assisting its own research
After ruling out simple explanations like increased compute, hiring surges, new data, and research breakthroughs, METR's best guess is that roughly 80% of the current acceleration comes from Anthropic using its own AI to assist research. But in the public report, most of the reasoning around this figure and the most persuasive evidence have been redacted by Anthropic; METR admits this forces outsiders to rely more on their judgment.
Researcher testimony aligns with end-to-end automation metrics
Internal Anthropic researchers said that in certain areas AI has dramatically accelerated work; one researcher's exact words are quoted. While this type of anecdotal evidence alone isn't compelling, it aligns with the timeline: the acceleration appears to have begun in October 2025, precisely when Claude could write code continuously and independently for hours. Anthropic's own data shows that earlier this year, Claude barely conducted research end-to-end, but by August could complete roughly a quarter of it under human supervision.
No clear factor will break this cycle in the near term
AI building AI differs from other technologies: the stronger the AI, the faster it can help you build the next generation. This cycle only gets stuck at three bottlenecks—compute, training data, or algorithmic insights. METR examined all three and found no signs of constraint within the next year; they admit to low confidence and changing views with new information, but cannot rule out further acceleration in the next six months, or even acceleration of the acceleration itself.
Toby Ord's challenge and METR's response
Philosopher Toby Ord pointed out that the test score line might have another explanation: Mythos comes from a much larger training run, and large training runs typically deliver a one-time capability jump. Excluding that factor, the acceleration is only 1.4x rather than 2x. But AI safety researcher Buck Shlegeris responded that test scores were never METR's main basis for their conclusion—they're only the part of the report that wasn't redacted and that the public could see. The evidence that made METR more confident in doubling was what got deleted.
Anthropic admits it has crossed its own safety threshold
Anthropic drew its own line: AI accelerating research to walk through two years of progress in one year. Its August public report claimed the acceleration was "below 2x," but using Anthropic's private data, METR calculated 1.97x. Just before the author finished this episode, Anthropic admitted in its latest IPO filing that based on expectations for the next generation of models and current velocity, the company has already crossed the line it drew for itself—but only said it is evaluating the impact, not what it will do.
Companies won't constrain themselves; regulation must step in
The logic for why AI companies won't stop is clear: if we stop, competitors will continue, and they might be even less careful. New York State legislator Alex Bores called for Congress to pass legislation banning AI companies from fully automating their own research; this proposal has supporters from Bernie Sanders to Steve Bannon. METR also points out that whether they get invited into the company, what they can see, and what the report says are all decided by Anthropic—the company even tightened its redaction rules in the final week, forcing METR to redo work and delay publication.
In their own words · checked verbatim
I’ve seen us discovering very promising directions entirely because people can spam 10 ideas at Claude and just see what sticks.
the line only shows progress about 1.4 times as fast
that means security strong enough to stop the world’s best-resourced hackers, including governments, and tracking everything that happens in its AI development
he says he came away convinced
by less than a factor of 2
Recursive self-improvement — AI building better AI building better AI — may be close.
Figures
| Anthropic's current research acceleration (METR estimate) | roughly 2x (confidence interval 1.3-3x) | 1:57 |
| Anthropic's acceleration calculated from private data | 1.97x | 10:41 |
| Acceleration attributed to AI-assisted research | roughly 80% | 4:00 |
| Claude's end-to-end research completion rate (August 2025) | roughly one-quarter | 4:48 |
| Claude's independent work duration (end of 2025) | over 30 hours | 2:57 |
| Math problem that puzzled mathematicians | 87 years | 2:57 |
| Toby Ord's acceleration estimate (excluding one-time gain) | roughly 1.4x | 7:33 |
| AI companies' probability of building superintelligence within one year (Buck Shlegeris estimate) | roughly one-third | 9:09 |
Glossary
- recursive self-improvement
- The cycle where AI assists in building stronger AI, repeating and accelerating continuously.
- AI 2027
- A widely cited predictive scenario report based on AI progress and geopolitical developments.
How to listen
Anyone focused on AI risk governance, working at an AI company or in policy, or considering a career change to AI safety.