METR Reveals Anthropic AI Research Acceleration Driven by AI Itself

Independent evaluator METR published an assessment on November 5, 2026 about AI research and development acceleration at Anthropic. The result of this assessment has attracted great attention in the AI world. Anthropic's internal AECI benchmark scores grew 1.9x faster since October 2025. The 95 percent confidence interval is between 1.5 and 2.8x.

Even more interesting, METR estimates AI tools contributed about 80 percent of the current acceleration. This is not due to added training compute, but because AI itself helps build newer models. In other words, AI is helping accelerate the development of the next generation of AI.

METR says it found no convincing bottleneck to continued rapid progress over the next 12 months. However, most supporting evidence is redacted by Anthropic. This raises questions about how much we can trust the presented data.

The Numbers Behind "1.9x Acceleration"

The number 1.9 means nothing before we know what was measured. AECI is Anthropic's internal benchmark for well-specified tasks, so a rising score means the model is getting better at finishing jobs that are already neatly defined. What changed is not that the model suddenly leapt in intelligence, but that the pace of improvement rose.

The 95 percent confidence interval between 1.5 and 2.8 times shows its honesty. The low end still means one-and-a-half-fold acceleration; the high end nearly triple. METR did not hide the width of that range, and that is exactly what makes the figure more trustworthy than a single round number.

One thing to underline: this does not measure humans losing a race. It measures AI tools shifting the capability curve upward. The 1.9x figure is a rate of capability improvement, not a count of tasks becoming easier. Those two things are often mixed up in coverage, and this article does not mix them.

How AI Helps AI: The Acceleration Mechanism

So how does AI accelerate AI development? The answer sits in three jobs that earlier models left almost entirely to humans.

First, writing code. AI coding assistants write and fix most of the infrastructure code in a lab before the human team touches it. If experiment code finishes twice as fast, the cycle of trying and measuring model capability also runs twice as fast. This is not about a model literally writing itself, but about speeding up one joint of the cycle.

Second, analyzing experiments. Newer models read experiment results, spot patterns humans miss, and propose the next experiment. The loop becomes: one generation helps direct the next, and then that generation helps again. A coffee-shop analogy works here: the helper used to be unable to say "today's coffee is short on grounds" — now he can, so the blend improves faster.

Third, compute orchestration. Scheduling GPU clusters, picking architectures, and tuning training parameters is heavy patterned work: AI can test thousands of combinations no human would have time to finish one by one.

That is what METR means when it credits AI tools with about 80 percent of the acceleration. It is not that compute grew by 80 percent, but that the work done by AI inside the research cycle did.

What Is AECI and Why It Matters

AECI is Anthropic's internal benchmark that measures model ability in performing well-specified tasks. METR audited Anthropic from August to October 2026. The audit goal is to determine whether AI is accelerating its own development.

This concept is known as Recursive Self-Improvement or RSI. It means AI can improve itself, then the better version improves the previous version again, and so on. If RSI is achieved, AI development could become very fast and hard to control.

Dario Amodei, CEO of Anthropic, previously wrote that AI may soon accelerate its research faster than researchers can evaluate it. METR provides data supporting that claim, though with redactions on some evidence.

Note that Amodei's claim comes from the most interested party: the CEO of the lab being measured. That is exactly where METR matters, because it is a third party that takes no payment from Anthropic and does not depend on the industry's profits rising or falling.

Deeper Dive: RSI Implications for AI Safety

AI research acceleration driven by AI itself has major implications. If development speed keeps increasing, regulators will struggle to keep up. Policies made today may be outdated tomorrow.

Here are several implications to consider:

Read the table from the bottom up and the pattern gets sharper: if human control is the first thing at risk, the other three risks are what follow it. A test that arrives too late is a safety problem; a rule that arrives too late is a governance problem. When the timeline compresses, verification and oversight compress with it, and the paperwork that used to separate them thins out.

The table above shows risks that need anticipation. METR suggests external evaluators be embedded inside AI labs. This idea is supported by Amodei and Sam Altman from OpenAI.

Implications for the AI industry are also real. If AI can accelerate its own development, AI companies will become increasingly difficult to challenge by new players. The cost to remain competitive will keep rising. Only companies with large resources can survive.

METR is a non-profit organization funded by donations. They do not accept compensation from the AI labs they evaluate. This gives a higher level of objectivity in their assessment.

Redacted Evidence: What Can and Cannot Be Verified

The weakest part of METR's report is also its most honest: most supporting evidence is redacted by Anthropic, the party being audited. The public sees METR's conclusions and only faint traces of figures that passed the redaction.

There are two ways to read this situation, and both make sense. The first: AI labs are entitled to protect trade secrets, and an evaluator granted privileged access has a legal duty to redact much of what it sees. The second: an audit whose detailed evidence the public cannot inspect asks for trust in the evaluator rather than data anyone can check.

Note that METR does not pretend to have unlimited access. It writes the limitation into its own report, complete with the wide confidence interval. That honesty about limits is not a weakness of the report; it is the reason the report can be read carefully.

What readers must not do is treat the 1.9x figure as if it were a factory-verified number. It is the best audit currently available, and even that redacts the supporting detail. A careful reader holds two thoughts at once: this is the best number available, and certainty about it remains limited.

Impact on Society

RSI development also has impact on society broadly. First, if AI can improve itself rapidly, automation will become more widespread in various sectors. This can change the form of jobs that exist. Second, flexible AI regulation is needed so that innovation is not hampered, but remains safe for society. Third, AI education needs to be expanded so that people understand how this technology works.

However, there are also concerns. If AI development is too fast without adequate supervision, security risks will increase. Governments need to monitor this development better and make policies that balance innovation and security.

There is a subtler impact: the public has long assumed that humans set the pace of AI progress. That assumption keeps public discussion comfortable, because there is always someone who can be held accountable and brought to the table. METR's report tells us that assumption may no longer be accurate, and most readers have not yet adjusted their expectations.

One thing to keep in mind: this report was published on November 5, 2026 by an evaluator paid by no lab, so its weight differs from a company statement. It does not prove RSI has arrived, only that its acceleration is measurable and its evidence is fragile. Both can be true at once, and both deserve to be read with the same care.

That is why this report matters beyond its 1.9 figure: it gives readers a reason to stop asking "when will AI change" and start asking "who is actually speeding up that change right now, and who is watching it".

Key Terms Glossary

• METR: Independent non-profit evaluator for AI research
• AECI: Anthropic's internal benchmark to measure model capability
• RSI: Recursive Self-Improvement, AI improving itself repeatedly
• Dario Amodei: CEO of Anthropic
• Sam Altman: CEO of OpenAI
• Confidence interval: The range within which the true figure most probably falls, not a single certain number