Anthropic's AI Is Getting Smarter at Making Itself Smarter — Twice as Fast as Before

Something significant is happening inside Anthropic's research labs, and the people tasked with watching for it are now saying so out loud. According to an independent evaluation published by METR, a safety-focused assessment organization, Anthropic's AI systems are now improving their own core capabilities at roughly twice the rate they were just a year and a half ago. That finding alone would be notable. The fact that most of the supporting evidence has been withheld from public view makes it considerably harder to know what to do with it.

What the Assessment Found

METR conducted a high-access evaluation at Anthropic between August and October 2026, with a specific mandate: determine how much artificial intelligence is accelerating Anthropic's own AI research and development. The result was a central estimate suggesting Anthropic is improving model capabilities on well-specified tasks at approximately twice the rate compared to the period spanning March 2024 through October 2025. The researchers were careful to note uncertainty in that figure, placing the plausible range between 1.3x and 3x acceleration.

The primary driver of this acceleration, according to METR's findings, is Anthropic's own extensive deployment of AI tools throughout its R&D pipeline. In other words, the AI is helping build a better version of itself — and it appears to be doing so with increasing efficiency. This recursive dynamic, where AI systems contribute meaningfully to the research that produces the next generation of AI systems, is something safety researchers have long flagged as a potential inflection point.

The Redaction Problem

Here is where the story becomes more complicated. While METR's high-level conclusions are publicly available through their published assessment, the majority of the underlying evidence has been redacted. Anthropic's stated rationale is that detailed technical information about how AI is accelerating their R&D could benefit competing AI developers — a competitive concern the company deemed significant enough to override public transparency.

That reasoning is understandable in a narrow business sense. It is also, frankly, deeply uncomfortable from a safety and oversight perspective. The entire value proposition of independent evaluation rests on the ability of outside observers — researchers, policymakers, civil society — to scrutinize the evidence and reach their own conclusions. When the evaluator's work is published but the substance is hidden, what remains is closer to an endorsement than an audit.

METR's researchers appear aware of this tension. Their published findings carry significant caveats about uncertainty, and the organization has made clear that the redactions constrain what any outside observer can conclude from the report.

Why This Moment Matters

The concern is not simply that Anthropic's models are getting better faster. The concern is what happens next, and whether anyone outside the company is positioned to see it coming.

METR's assessment noted the absence of observed bottlenecks or strong diminishing returns in the current trajectory. That is a technically precise way of saying there is no obvious reason for the pace of improvement to slow down. In a system where acceleration is compounding — where AI-assisted R&D produces better AI, which further assists R&D — the absence of natural brakes is precisely the scenario that safety researchers have been warning about for years.

Rapid capability gains, particularly those driven by recursive AI involvement in its own development, raise the possibility of sudden capability emergence: AI systems crossing meaningful thresholds quickly, without the gradual lead time that would allow for deliberate safety responses. This is the scenario that makes the redaction issue more than a transparency complaint. It is a question of whether the oversight mechanisms that are supposed to protect against dangerous AI development can function when the evidence underpinning them is treated as proprietary.

The Bigger Picture

Anthropic has consistently positioned itself as the safety-conscious actor in an industry not always known for caution. That reputation makes the tension here sharper, not softer. Commissioning an independent evaluation is a meaningful step. Redacting most of it tests the credibility of the exercise.

As AI systems become more deeply embedded in their own development cycles, the pressure on external evaluation processes will only grow. METR's assessment is a data point. Whether it becomes a turning point depends on whether the broader ecosystem — companies, regulators, and the public — treats acceleration of this kind as something requiring real accountability, not just a footnote in a partially visible report.