Anthropic’s artificial intelligence is now improving about twice as fast as it did a year ago, according to an independent assessment published on Thursday by the nonprofit evaluator METR. The group spent eleven weeks inside Anthropic examining internal data, benchmarks and staff interviews, and concluded the lab has packed roughly two years of historical progress into the past year. METR attributes most of the speedup to Anthropic’s own AI systems doing research work, a finding that lands as the company prepares one of the [largest public listings in history](https://geroundtable.com/anthropic-ipo-amazon-biggest-winner/).
## Two years of progress in one year
The number comes from AECI, Anthropic’s version of the Epoch Capabilities Index, which rolls many benchmark scores into a single measure. Since October 2025 that index has grown about 1.9 times faster than its earlier pace, with a 95 percent confidence range of 1.5 to 2.8 times. The trend steepened from 15 points a year to 29 points a year.
METR stresses the uncertainty around that figure. Its judgment is that the true acceleration could sit anywhere from 1.3 times to 3 times the historical rate, and benchmark scores can mislead when models are tuned to the tests. Even so, the group says the pattern held across different ways of cutting the data, with estimates from 1.1 to 2.9 times.
## AI doing the research
The more striking claim is why the pace changed. METR estimates that roughly 80 percent of the current acceleration comes from Anthropic using its own AI to do research and development, though it calls that figure highly uncertain. Researchers told the group their work runs about twice as fast with AI help.
That help is mostly assistive rather than autonomous, METR says. AI systems write code, run experiments and draft analyses, but they are not yet making independent contributions to the most important stages of training models. The report adds there are signs autonomous contributions will play a larger role soon, and staff self-reports of a twelvefold speedup kept climbing through October.
## The evidence the public cannot see
There is a catch that runs through the whole report. Anthropic had the right to redact any non-public information, and METR says the company removed a majority of the evidence it found most informative. Anthropic described the redactions as aimed at keeping confidential material from accelerating rival developers. The result is that readers must largely take METR’s judgment on trust.
METR is blunt about that limit. It writes that the public cannot judge the redacted evidence or reconcile it with other sources, and that its conclusions rest on judgment more than would be ideal. The group notes that Anthropic changed the scope midway, dropping a planned check on how much of its computing goes to capabilities work versus safety work.
## A threshold that may already be crossed
The assessment reviewed the automated research section of [Anthropic’s August risk report](https://geroundtable.com/anthropic-s1-training-pause-model-escapes/). METR agreed the risk from models as of mid-July was low, but questioned the company’s claim that it had not crossed its own early-warning line, a doubling of the pace of AI research. METR calls that judgment contestable and says it is plausible the line was already crossed.
That sits awkwardly against the company’s own public stance. In September, chief executive Dario Amodei published an essay calling on the industry to slow advanced AI development so safety measures can catch up. Anthropic did not immediately respond to a request for comment on the METR findings.
## Our read
The headline number will get the attention, but the redaction story is the more important one. An independent check that cannot show most of its best evidence is a strange kind of transparency, and METR knows it. The group still published a clear verdict rather than hiding behind caveats, which gives the finding weight even while outsiders cannot test it.
What matters is the direction, not the decimal. If AI systems are doing a large share of the research that builds their successors, the labs with the [best internal tools](https://geroundtable.com/claude-writes-90-percent-anthropic-code/) pull away faster than rivals can copy them. That compounding advantage is why the report’s warning about no bottlenecks should worry anyone counting on the pace to slow by itself.
## Why It Matters
This is the first independent measurement of the idea the industry has debated for years: AI building better AI. A doubling of the improvement rate changes the timeline for every safety plan, investment thesis and regulation written on the assumption that progress moves at its old speed. It also arrives days before Anthropic asks public markets to value that acceleration.
## Background
METR, short for Model Evaluation and Threat Research, is a nonprofit that evaluates frontier AI systems and has run capability assessments for Anthropic, OpenAI and other labs for years. Anthropic lists AI research acceleration as a key risk indicator in its Responsible Scaling Policy, and its chief executive has argued the industry should slow down while safety work catches up.
Source: [METR’s assessment](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/) and [the AP report](https://techxplore.com/news/2026-11-anthropic-rapidly-ai-slowdown-group.html)
0 Comments