Anthropic spots early signs of AI self-improvement in Model 2

AI models might still be inching forward, but the rulers we use to measure them are starting to break down.

In its August risk report, Anthropic shared an update on Model 2, which has already become the lab's primary internal workhorse. While Model 2 edges out Mythos 5, the team admits the jump is merely an incremental improvement rather than the dramatic leap seen between Opus 4.6 and Mythos Preview. As a result, Anthropic determined it still falls below its RSP safety thresholds.

Routine anti-distillation and broken benchmarks

The most revealing details sit outside raw capabilities. Anthropic disclosed that anti-distillation has become a permanent, routine engineering investment to keep rivals from siphoning outputs.

Meanwhile, standard task-based benchmarks are losing measurement precision, making it hard to track just how much smarter models are getting. Yet in the realm of recursive self-improvement (RSI), Anthropic claims it has spotted early signs of acceleration, even while keeping the specific data under wraps.

Why it matters

This update highlights two major shifts at the frontier of AI.

First, conventional generational leaps appear to be narrowing, forcing leading labs to burn serious engineering bandwidth on defensive moats like anti-distillation.

Second, the critical bottleneck is pivoting from training to evaluation. If existing benchmarks can no longer capture the ceiling of new models, tracking the quiet beginnings of self-improvement will require entirely new measurement tools before things actually spin out of control.