Gemini 3.1 Pro Preview vs Kimi K3

Independent runs hand Kimi K3 one of the three agent-work benchmarks outright: Toolathlon-Verified 76.5 vs 61.1, a real gap with every number independently run. Agents' Last Exam (28.3 vs 16.4 overall pass rate) was a second until 2026-10-02, but Kimi's row is a Max run and Gemini's a High run, so it is setup-dependent rather than a win; the third, AA-AnalystAgent (added 2026-09-29), is a tie at 41.25 vs 38.75 pass^5, both Artificial Analysis runs, well inside the 11.2-point band. (An earlier version of this page said Gemini led Agents' Last Exam; that was our metric mix-up — partial-credit score vs pass rate — corrected 2026-08-17.) Terminal-Bench 2.1 now has an independent run on both sides, but from different harnesses — vals.ai's 80.9 for Kimi against tbench.ai's 65.8 for Gemini — so it is setup-dependent, not a win; on vals.ai's own board, which ran both, Kimi K3 leads its "Gemini 3.1 Pro Preview (02/26)" row 80.90 to 70.79, 10.1 points, just inside the 10.6 band. (Kimi's self-reported 88.3 was replaced by the vals.ai run on 2026-10-01.) Gemini 3.1 Pro ties on HLE no-tools and costs less ($2/$12 vs $3/$15); LiveCodeBench (88.48 vs 87.19) would be a tie too, but since 2026-09-29 that benchmark is graded saturated here and no longer called. LiveBench is a tie too (77.0 vs 79.2); ARC-AGI-2 doesn't compare cleanly here either — Gemini's tracked score is untiered (77.1) while Kimi K3's is its Max tier (60.4), no matched-tier pair to call. HMMT Feb 2026 is tainted, not a tie: both scores (94.70 vs 96.97) carry MathArena's own disclosed contamination-warning, since every tracked model postdates the competition, so this site excludes the whole benchmark from real/tie calls rather than reading the 2.27-point gap as meaningful. GPQA Diamond is out of the count now too — it's graded saturated (grade D) as of this review, so the 95.45 vs 92.93 gap here is ceiling compression, not a real tie.