How Far Are We from AGI When Agents Learn to "Think" Instead of "Tune Parameters"

The company CiYuanWuXian (Infinite Tokens) has raised three funding rounds in just one year, led by a former ByteDance female executive, and its product outperforms GPT-5 on coding benchmarks. I read this news several times in the lab—not because of the fundraising speed, but because of the result itself: "outperforming GPT-5." As a first-year master's student still grinding through papers, my biggest confusion is: How can a company founded less than a year ago surpass top-tier models like GPT-5 and Claude on tasks like coding, which demand extremely high logical rigor?