Anthropic uses existing AI models to train future AI models at a dizzying pace. It also stocks new models.
That’s according to a new report from third-party security researchers released Thursday — one that researchers say was heavily redacted at Anthropic’s insistence.
“While we were able to freely share our highest risk conclusions, Anthropic suppressed the majority of evidence that we found most informative for these conclusions,” METR said in its report.
In writing their report, researchers from METR, a research group perhaps best known for its report on the OpenAI Hugging Face agent breach, said they received “unprecedented access” to Anthropic’s internal data for nearly three months.
The group of researchers interviewed employees, surveyed Slack channels, and examined data on previously unseen internal AI models.
Here are three takeaways from their report.
Anthropic recently completed 3 new models that it has yet to release
Anthropogenic workers use better AI models than the rest of us.
In fact, METR’s report states that over the past six months, the company has completed training on three models that are not yet publicly available.
METR found that Anthropic’s internal-only models include one from 2022, as well as models completed in July, August or September, and September or October.
These internal-only models help Anthropic improve its AI capabilities even faster, according to METR estimates. This disclosure highlights one of the AI industry’s high-profile acts.
Anthropic and OpenAI make money by selling the most advanced AI models to the public. However, keeping cutting-edge models private also gives companies’ product and research teams a competitive advantage.
Executives could keep models private longer for security reasons. Anthropic and OpenAI’s batches of high-performing internal models are part of the reason METR Ajeya Cotra told Business Insider earlier this year that she expects the “most serious incidents” in AI security to occur at both companies.
Anthropic staff say their productivity is soaring
Anthropic has revealed data from an internal Slack survey showing that the AI giant’s employees are reporting significant productivity gains.
METR wrote that the average self-reported increase was about 12 times “for the best performing internal model” in October. METR cautioned that it considers the Slack surveys to be weak evidence because they may not represent all Anthropic employees. The research group also cautioned that respondents may not have given much thought to their answers.
When METR asked Anthropic researchers about their experiences, many talked about how Claude had significantly boosted their work.
“Now there are just bottlenecks on ideas and calculation, more on code,” one researcher told METR, agreeing with the idea that Claude has at least doubled their productivity.
What is more striking is that the METR did not see any “bottlenecks” that would prevent the acceleration from accelerating further over the following year.
“The absence of observed bottlenecks or sharply diminishing returns means that there is a significant risk that further advances in capabilities will lead to even faster acceleration, although, as we argue below, we risk missing important bottlenecks,” METR writes.
Recursive self-improvement fuels rapid development
Anthropogenic engineers are increasingly delegating monitoring and training responsibilities to other AI models, a risky process known as recursive self-improvement.
The method pays off in terms of speed.
METR found that Anthropic made “approximately two years of historic progress over the last year.”
The researchers said they were told that the acceleration of research and development was being achieved “primarily through AI assistance rather than fully autonomous contributions.”
“But there are signs that fully autonomous contributions will soon play a larger role,” METR said.
Jacob Coxon, the former Anthropic researcher who left the company in September, highlighted recursive self-improvement as a factor that could cause AI systems to spin out of control.
“They are rushing straight to self-improving superintelligence and playing with our lives,” he said of OpenAI and Anthropic.
Correction: November 5, 2026 — This story has been updated to clarify the type of access researchers received and the number of takeaways from the report.
Gn bussni
