METR, a nonprofit that measures AI capabilities, released research showing Anthropic's Mythos model significantly exceeds existing benchmarks for long-horizon AI agent reliability. At 50% success rate, Mythos breaks through METR's 16-hour measurement ceiling—the longest duration their current task suite can reliably test. More critically, at 80% reliability (the standard for practical deployment), Mythos handles tasks requiring humans over three hours, substantially surpassing Gemini 3.1 Pro, the closest competitor measured. The breakthrough reveals a key misconception: the 16-hour metric measures task complexity/difficulty, not actual runtime. As base models become more capable, agents can sustain focus on complex goals for longer before context degradation becomes fatal. This suggests autonomous AI systems capable of independent work over extended periods are arriving faster than many expected, with major implications for developers building agent-based products.