Over Eight Months of Progress in Two: Analyzing the Mythos Preview Capability Jump
Capability growth rate quadruples, with time horizon exceeding 87 hours.
Anthropic's latest model, Claude Mythos Preview, has alarmed AI forecasters with its sudden capability jump. According to analyses by Alvin Ånestrand on LessWrong, the model's Epoch Capabilities Index (ECI) score rose from ~153 (Opus 4.6 in Feb 2026) to ~161 (Mythos Preview in Apr 2026) in just two months. Previously, such gains took roughly six to eight months of steady progress. The speedup ratio ranges from 1.86× to 4.28× depending on the breakpoint chosen, implying a possible doubling of AI capability growth rates.
This acceleration translates to startling real-world impacts. The model's 50% time horizon—the task length it can complete with 50% success—jumped from ~12 hours for Opus 4.6 to an estimated 87 hours for Mythos Preview, meaning it can autonomously handle work that would take a human over two full work weeks. In dedicated AI R&D tests, Mythos Preview succeeded at two out of three tasks requiring more than 40 hours of human expert effort. If this pace continues, the time horizon could double every month, signaling the onset of recursive self-improvement and potentially an intelligence explosion.
- ECI score jumped 8 points in 2 months (Opus 4.6 at ~153 to Mythos Preview at ~161), previously requiring 6-8 months.
- Capability growth speedup estimated between 1.86× and 4.28×, with a middle estimate of ~3× faster progress.
- 50% time horizon estimated at 87 hours, up from 12 hours, enabling completion of multi-week human tasks.
Why It Matters
If sustained, this acceleration could trigger recursive self-improvement, compressing years of AI progress into months.