The second model is biased, so $x_1$'s effect is underestimated

$x_2$ is confounding or sharing variance with $x_1$; the new coefficient represents $x_1$'s unique effect

The increase in $R^2$ indicates $x_1$'s true effect is actually 2.8

The first model was missing data, so 4.0 was incorrect