“Cognition's internal evaluations show that AI agents consistently achieve better scores when allowed to use multiple models versus being constrained to a single model.”