Which is better, Gemini 3.1 Pro or Kimi K2.6?

Across 94 shared benchmarks tracked on Vector Wire, Gemini 3.1 Pro leads on 61 and Kimi K2.6 leads on 30 — Gemini 3.1 Pro leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

Gemini 3.1 Pro vs Kimi K2.6

GooglevsMoonshot94 shared benchmarks61–30 head-to-head

Benchmark	Gemini 3.1 Pro	Kimi K2.6
AA Agentic Index	23	58.7
AA Intelligence	47.7	45.1
AA-LCR	79	76.7
AA-Omniscience	32.9	6.4
AgentWorldBench - Android	61.4	58.9
AgentWorldBench - MCP	59.1	65.2
AgentWorldBench - OS	66.9	60.8
AgentWorldBench - Overall	54.6	53.4
AgentWorldBench - Search	30.2	27.5
AgentWorldBench - SWE	59.1	58.8
AgentWorldBench - Terminal	52.5	52.5
AgentWorldBench - Web	52.8	50.2
aime_2026	98.3	96.4
Apex (Pass@1)	60.9	24
Apex Shortlist (Pass@1)	89.1	75.5
apex_agents	33.5	27.9
apexAgents	33.5	28.5
arena_elo	1486	1461
arena_text_factuality	1471	1455
arena_vision	1280	1264
Artificial Analysis Coding Index	68.8	61.8
baby_vision	51.6	68.5
baby_vision_with_python	68.3	68.5
browsecomp	85.9	86.3
BrowseComp (Agent Swarm)	85.9	86.3
BrowseComp (w/ Ctx)	85.9	83.2
charxiv_rq	89.9	86.7
charxiv_rq_with_python	89.9	86.7
Chinese SimpleQA (C-SimpleQA)	85.9	75.9
coding_arena_elo	1447	1509
critpt	17.7	8
deepsearchqa_accuracy	60.2	83
deepsearchqa_f1	81.9	92.5
Finance Agent v2	43	44.9
FORTRESS (Adversarial)	65.2	65.6
FORTRESS (Benign)	98	97.2
frontiermath_tier_4	16.7	14.6
gdpval	23.3	41.4
GDPval-AA (Elo)	1317	1482
GDPval-AA v2	962	1190
Global-MMLU-Lite	92.7	88.4
GPQA Diamond	94.3	91.1
HiL-Bench	35.3	18.7
HLE	51.4	37.5
HLE (with tools)	51.6	54
hmmt_feb_2026	94.7	94.7
IFBench	77.1	76
imo_answer_bench	91	86
itbenchSre	30.3	31.2
livebench	79.9	72.2
livebench_agentic_coding	44.1	46.9
livebench_coding	76.5	78.6
livebench_data_analysis	78.5	65.1
livebench_instruction_following	79.1	64.4
livebench_language	85.4	75.1
livebench_math	91	84.3
livebench_reasoning	84	79.4
LiveCodeBench	91.7	89.6
LiveCodeBench v6	91.7	89.6
mathvision	89.8	93.2
mathvision_with_python	95.7	93.2
MCP Atlas	78.2	68.1
MCPAtlas Public (Pass@1)	69.2	66.6
mcpmark	55.9	55.9
MMLU-Pro	91	87.1
mmmu_pro_with_python	85.3	80.1
MMMU-Pro	83	80.1
OJBench (python)	70.7	60.6
OmniScience Accuracy	55.3	32.8
OmniScience Non-Hallucination	50.1	60.7
OSWorld-Verified	76.2	73.1
Prophet Arena	0.1	0.1
scicode	59	53.5
simpleqa_verified	77.3	38.7
strongreject	98	99.8
SWE-bench Multilingual	76.9	76.7
SWE-bench Pro	54.2	58.6
SWE-bench Verified	80.6	80.2
SWEBench Pro (Public)	54.2	58.6
Tau 3 Banking	16.5	20.6
TauBench V3 - Banking	21.4	23.3
Terminal Bench 2.1 (Best Harness)	73.8	71.3
Terminal-Bench 2.0	68.5	66.7
Terminal-Bench 2.1	74	65.9
Terminal-Bench Hard	68.5	43.9
toolathlon	48.8	50
usamo_2026	74.4	51.2
V* (w/ python)	96.9	96.9
vectara_answer_rate	99.4	99.7
vectara_avg_summary_length	107.7	116.7
vectara_factual_consistency	89.6	89.2
vectara_hallucination_rate ↓	10.4	10.8
τ²-Bench Telecom (AA run)	99.3	95.9
τ³-Bench	67.1	20.6

Best tracked score per model per benchmark (default configuration; source-attributed). ↓ marks lower-is-better metrics. Open either model for its full surface, provenance and pricing.