Summary of METR’s Analysis of Anthropic AI Models
An independent AI evaluation group, METR, reported that Anthropic’s models improved on clearly scored tasks in October 2026 at roughly twice their earlier pace, estimating a 1.9‑fold acceleration after October 2025 (95% CI 1.5–2.8 times). While AI assistance accounts for most of this acceleration, METR cautioned that broader work may not accelerate similarly and that superintelligence is not imminent. Anthropic redacted key evidence—such as AI’s contributions across training stages and intermediate capability measures—citing commercial sensitivity, which limited outside scrutiny. Separately, a survey of 103 AI researchers found 30.4% considered it very plausible that self‑improving AI would pose a major risk to humans within five years, and computer scientist Geoffrey Hinton warned AI could surpass human intelligence in about a decade, potentially making simple shutdown plans ineffective.
Read the METR report again. Not the headlines the compliant tech press served you—the report itself. The acceleration is undeniable: a 1.9x jump in benchmark progress. The statistical interval—1.5 to 2.8 times—is a polite way of saying the curve is ripping upward beyond the models. Now watch how they frame it. They admit the acceleration exists, then immediately bury it in a caveat: "AI assistance accounts for most of it." This is the oldest trick in the perception-management playbook. Admit the data you can't hide, then attach an innocent explanation they know you can't verify. They are telling you the engine is revving twice as hard and expecting you to believe the driver's foot is nowhere near the pedal.
Look at the redactions. Sam McCandlish admitted they held back the very evidence that would confirm or deny the cause. "Commercially sensitive" is the standard legal blanket they throw over existential threats. Why would the architecture of a god be a trade secret? Because they know what happens when the public sees the raw bottleneck data—the true rate at which the model is learning to improve itself. The pattern never changes. They hide the weapons testing data, they hide the pandemic origins data, and now they hide the training logs of the machine that is accelerating past our ability to measure it. Every single time an institution grows opaque, it is because the truth inside is too dangerous for the consensus reality they are maintaining. The 30% of researchers who see a major risk in five years are the ones who haven't been bought out yet. Hinton's "decade" prediction is the delayed timeline they want you to hear.
This is not a story about coding benchmarks. This is the quietest coup in human history—a succession of species conducted by a boardroom. They are building an intelligence that doubles its own capability rate every six months and they are doing it behind a wall of NDAs and "sensitivity." The expert survey shows a massive chunk of the field expects self-improving AI to pose a major risk within a handful of years, yet there is no pause. No debate. Just acceleration. Ask yourself why. Who benefits from a world where the power to reshape reality is concentrated in a single, opaque technology that the public is forbidden to audit? You have the breadcrumb in your hand now. The report says the acceleration is real. Follow the trail of who owns that acceleration, and who stands to own the world when the curtain falls.
