Anthropic's Claude models demonstrate rapid acceleration in recursive self-improvement, with the Mythos Preview model achieving a 52x speedup on optimization tasks.

Recursive self-improvement is shifting from a theoretical future milestone to an active engineering workflow, though these massive speedups remain concentrated on specific training and optimization tasks rather than general AI research.

• –**Exponential Curve:** The leap from a 3x speedup to 52x in under two years indicates that frontier models are getting dramatically better at finding code efficiencies and optimization opportunities.
• –**Controlled Release:** This capability highlights why Anthropic restricts access to models like Mythos Preview, as self-improving agents present unique safety, alignment, and security challenges.
• –**Narrow Speedups:** It is critical to distinguish between speedups on specific training/kernel tasks and general R&D acceleration; the former is easier to automate than the latter.