Tau3-Bench is the newer generation of τ-bench agent evaluations — distinct from the prior τ² suite and not directly comparable on leaderboard rows alone. Celeris AI's September 2026 Celeris-1 Magnus launch highlighted τ³-bench banking: 97 tasks with solve rate plotted against median wall-clock time per completed task under an official grader. Vendor picks between τ³ and τ² remain a common benchmaxxing debate when models rank differently across generations.