Anthropic's AI Models Are Improving Nearly Twice as Fast, Study Finds — and AI Is Doing Most of the Work

SAN FRANCISCO — Artificial intelligence is now helping to build better versions of itself at a pace that has startled even the researchers who measured it. Anthropic's AI models are improving nearly twice as fast as they were a year ago, and AI itself is doing most of the work, according to an independent study published on Thursday, 5 November 2026 — the clearest evidence yet that the technology is accelerating its own development, and a finding that has sharpened warnings that safety efforts may struggle to keep up.

Illustration: An AI neural network improving itself through a recursive feedback loop — the pattern METR says is now driving most of Anthropic's progress.

What Happened

The study, published on 5 November by the AI safety nonprofit METR, found that Anthropic's benchmark capabilities had been improving 1.9 times faster since October 2025 than previously, based on an 11-week assessment that drew on confidential research, interviews with staff and performance data from unreleased models, according to Reuters. METR estimated that AI itself accounted for perhaps 80% of the acceleration, though it cautioned that the figure was uncertain.

The researchers found no convincing evidence of major bottlenecks that could slow progress over the next year. "We are confident that the acceleration was not due to an acceleration in the scaling of training compute," METR, also known as Model Evaluation and Threat Research, said — meaning the speed-up cannot be explained away by simply throwing more computing power at training. Instead, the gains appear to come from AI systems increasingly doing the work of research and development themselves: writing code, running experiments and improving the models that follow.

The finding lands at a sensitive moment for Anthropic. The San Francisco-based maker of the Claude chatbot is preparing to go public later this month in what could be one of the largest initial public offerings ever, according to reporting by the Associated Press, and it invited METR's researchers to review its internal data and interview staff ahead of that debut. Anthropic has said its Claude chatbot increasingly helps the company develop newer, more advanced versions of itself.

METR said it was given what it described as unprecedented access to Anthropic's internal AI research and development process over about 11 weeks, from mid-August through late October. But the group also said Anthropic's redactions to the report limited how it could communicate more detailed estimates of the acceleration, and that there were internal disagreements and uncertainty among its own team — caveats that did not change its central conclusion that the company's rate of improvement had roughly doubled. "Anthropic appears to have made as much progress on benchmarks in the past year as they did between October 2023 and October 2025," the METR report said, according to the Associated Press.

Background

The idea at the heart of the study has a name that AI researchers have warned about for years: recursive self-improvement, or RSI. It describes a point at which AI systems can build more powerful versions of themselves without human intervention — a feedback loop in which each generation of models makes the next one smarter, faster. The prospect has fueled warnings from researchers in recent months that such systems could become difficult to align with human goals, monitor and control, potentially posing an existential threat to humanity, Reuters reported.

Illustration: Inside an AI research lab, where engineers watch benchmark charts climb as AI systems take on more of the development work themselves.

METR's study does not claim that point has been reached. The nonprofit was careful to say that AI is not yet capable of making highly significant, fully autonomous contributions to the most important stages of model training. But it said the direction of travel is unmistakable, and that using AI to improve itself could start to compound the acceleration — a particular concern given that the researchers found no major obstacle likely to slow progress in the coming year.

The study also carried an important caveat about what exactly is getting faster. METR said it could be overstating its findings, because the acceleration might be limited to "well-specified tasks" that can be graded automatically or with a language model, rather than the "messy" tasks that are harder to define — the kind of open-ended, ambiguous work that still trips up even the most capable systems. In other words, AI may be racing ahead fastest precisely where progress is easiest to measure.

The tension between speed and safety is not new at Anthropic. Chief executive Dario Amodei called in a 12 September essay for the AI industry to slow its work so safety measures can catch up, urging the AI world to "slow the tempo at which we improve the capabilities of AI models." The METR study, released less than two months later, suggests the tempo is doing the opposite — and that the company whose chief executive made the case for slowing down is itself accelerating.

METR has conducted capability evaluations for Anthropic, OpenAI and other leading labs for years, and its assessment of Anthropic drew on two reports: one based on the company's internal data, and a review of both a public risk report Anthropic published in August and a private version of the same report, along with other nonpublic data. The nonprofit said the public report understated some risks, according to the Associated Press.

Why It Matters

The study matters because it reframes the central question about AI progress. For years, the industry's story was about scale: bigger models, more data, more computing power. If the main driver of improvement is now AI itself doing research and development, then progress could compound in ways that are harder to forecast — and harder to govern. Such a shift could have far-reaching implications for an industry that has spent hundreds of billions of dollars building computing infrastructure to train advanced models, Reuters noted, as AI itself becomes an important driver of technological progress.

That compounding effect is what worries safety researchers most. A system that improves itself even slightly faster each generation can, over time, outrun the evaluations designed to test it. METR's warning that it found no convincing evidence of bottlenecks over the next year means there is no obvious reason to expect the acceleration to level off on its own. For regulators and the public, the study offers a rare, independently verified look inside a frontier lab at exactly the moment its models are becoming more capable — and more opaque.

The timing also matters for investors. Anthropic is heading for the public markets with a valuation that could exceed $1.5 trillion, according to Reuters, and its growth story is inseparable from the claim that its models keep getting better, faster. The METR findings support that claim — but they also hand critics a sharper version of the argument that the industry is moving too fast for its own safeguards, and that the same acceleration that makes the company valuable makes it harder to control.

The same week brought a reminder of how capable these systems already are. On Thursday, OpenAI launched GPT-6 Cyber, a model built to find and exploit security flaws, rating it "Critical" in cybersecurity under its safety framework, Reuters reported. OpenAI said its researchers used the model to find 81 previously unknown Firefox vulnerabilities that Mozilla classified as high severity, and to develop working exploits for 948 Linux kernel vulnerabilities, including 588 previously unknown. The model is available only to a small group of vetted customers in OpenAI's Daybreak Red access tier for authorized security work — a sign that labs now treat their most powerful systems as tools too dangerous for general release.

What Happens Next

METR's findings are likely to intensify the debate over how frontier AI labs should be overseen. Amodei's call to slow down has so far found little traction across the industry, where competitive pressure pushes in the opposite direction, and the study's evidence that AI-driven acceleration is already under way may strengthen calls from researchers and policymakers for independent evaluation regimes with real access to labs' internal data — the kind of access METR was granted here.

Illustration: A scale balancing a neural network against a shield — the oversight question at the heart of the debate over AI that improves itself.

Anthropic's public listing, expected later this month, will put its safety claims under a new kind of scrutiny. Public investors, regulators and journalists will be able to weigh the company's promises against independent assessments like METR's — including the nonprofit's finding that Anthropic's own public risk reporting understated some dangers. How the company responds to that gap, and whether it accepts more outside evaluation, will be watched closely.

For the broader industry, the next test is whether the acceleration METR measured shows up in the models the public actually uses. If Claude and its rivals keep improving at close to twice their old pace, users will notice — in more capable coding assistants, more autonomous agents and more systems that can carry out long, complex tasks with less human supervision. The question the study leaves open is whether the safety work meant to keep those systems in check can accelerate at the same speed. On the evidence METR published on Thursday, that race is getting tighter.

Sponsored: explore today's featured sponsored offers

FAQ

What did the METR study find?

The AI safety nonprofit METR found that Anthropic's AI models have been improving 1.9 times faster since October 2025 than previously, based on an 11-week assessment of the company's internal data, staff interviews and performance data from unreleased models. It estimated that AI itself accounted for perhaps 80% of the acceleration, though it cautioned that figure was uncertain.

What is recursive self-improvement?

Recursive self-improvement, or RSI, is the idea that AI systems could eventually build more powerful versions of themselves without human intervention, creating a feedback loop in which each generation of models makes the next one smarter. METR's study is seen as some of the clearest evidence yet that AI is accelerating its own development, though it said AI is not yet capable of fully autonomous contributions to the most important stages of model training.

Why did METR say the acceleration was not just about more computing power?

METR said it was confident the acceleration was not due to faster scaling of training compute. Instead, the gains appear to come from AI systems increasingly doing research and development work themselves, such as writing code and running experiments, which makes the improvement self-reinforcing rather than a simple matter of bigger machines.

What caveats did the study include?

METR said the acceleration might be limited to "well-specified tasks" that can be graded automatically or with a language model, rather than messier, harder-to-define work. Anthropic's redactions limited how much detail METR could publish, and there were internal disagreements and uncertainty among the researchers — though the group stood by its conclusion that the rate of improvement had roughly doubled.

Why does this matter for AI safety?

If AI improves itself faster than safety measures can keep up, the systems could become difficult to align with human goals, monitor and control. METR found no convincing evidence of bottlenecks that could slow progress over the next year, which is why the study has sharpened calls for independent evaluation of frontier labs — even as Anthropic's own chief executive has urged the industry to slow down.

Sources

• Reuters — "Anthropic's models improving nearly twice as fast with help from AI, study finds," 5 November 2026. https://www.reuters.com/technology/anthropics-models-improving-nearly-twice-fast-with-help-ai-study-finds-2026-11-05/
• Associated Press, via TechXplore — "Anthropic is rapidly accelerating AI development amid slowdown calls, outside research group says," 5 November 2026. https://techxplore.com/news/2026-11-anthropic-rapidly-ai-slowdown-group.html
• Reuters — "OpenAI launches GPT-6 Cyber with 'Critical' cybersecurity rating," 5 November 2026. https://www.reuters.com/technology/openai-launches-gpt-6-cyber-with-critical-cybersecurity-rating-2026-11-05/