Anthropic AI R&D Progress Hits 2x Pace

Anthropic’s biggest breakthrough may be its faster research loop, not a public model release. The evidence is significant, but it signals accelerated iteration—not confirmed recursive self-improvement.

• –AECI growth accelerated 1.9×, with a wide 1.5–2.8× confidence interval.
• –Model 4 clearing 50% across every tested benchmark suggests the evaluation suite is nearing saturation.
• –METR estimates roughly 80% of current acceleration comes from AI use, though self-reports and redacted evidence limit causal certainty.
• –Training-compute growth was not the driver, implying gains may come from better algorithms, tooling, and AI-assisted research workflows.
• –Developers should expect faster frontier-model and coding-agent iteration, alongside greater pressure for independent evaluations.