METR says Anthropic’s AI progress has doubled, mostly fueled by its own models

Anthropic has long built its brand on being the responsible grown-up in the room, preaching safety and deliberate pacing. Behind closed doors, though, the speedometer tells a very different story.

AI-evaluation nonprofit METR just released a landmark report after spending 11 weeks inside Anthropic, poring over internal data and interviewing staff. The headline takeaway is staggering: since October 2025, Anthropic’s internal model development has sped up by roughly 1.9x, with METR’s broader estimate sitting between 1.3x and 3x.

The real kicker is what is driving the sprint. METR attributes around 80% of this acceleration directly to Anthropic’s own models helping with internal R&D. Basically, the models are now doing the heavy lifting to build their own successors.

Opus 5.5 gets demoted

The report also gave outsiders a rare peek at Anthropic’s unreleased pipeline. In one eye-opening chart, Claude Opus 5.5 — currently regarded as a premier public model — was casually categorized as "non-frontier."

Sitting comfortably above it are internal experimental systems dubbed "Model 3" and "Model 4." Model 4 in particular demonstrated huge leaps over a remarkably short timeline, clearing the 50% mark across all 15 evaluated benchmarks.

Even seasoned evaluators were caught off guard. METR’s Ryan Greenblatt admitted he was "seriously wrong" after anticipating a much slower rate of progress this year.

Yet if you are hoping for a technical blueprint, you are out of luck. Anthropic heavily redacted the strongest pieces of evidence in the published version. Co-founder and safety lead Sam McCandlish explained the decision frankly, stating the lab wanted to avoid "handing out a roadmap for others to go faster."

There is also a governance catch. In its IPO filing, Anthropic specifically flagged "dramatically accelerated progress attributable to AI" as a major hazard threshold. With METR observing no obvious technical bottlenecks slowing the loop down, the lab appears to be inching uncomfortably close to its own danger line.

Why it matters

For builders and end users, Model 4 won't hit your API dashboard tomorrow, but it confirms that model lifecycles are shrinking fast. Capabilities once expected in 2027 could land well ahead of schedule, while today's flagship models will become yesterday's commodities at an accelerating clip.

For the AI industry, this is empirical confirmation that the recursive self-improvement loop has officially begun. When AI contributes 80% of research velocity, the competitive moat shifts from raw human talent to who can build the most effective automated experimentation workflows. Competitors now have no choice but to let their own models take the wheel in R&D.

We think Anthropic's stance reflects an intense internal paradox. Redacting the technical roadmap makes sense as a defensive business maneuver. But watching an AI lab run toward the exact hazard threshold it publicly warned about shows just how irresistible the flywheel is. When AI starts accelerating AI with no clear ceiling in sight, slowing down is no longer an option anyone can afford.