Which is better, Kimi K2.5 or Kimi K2.6?

Across 69 shared benchmarks tracked on Vector Wire, Kimi K2.5 leads on 4 and Kimi K2.6 leads on 63 — Kimi K2.6 leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

Kimi K2.5 vs Kimi K2.6

MoonshotvsMoonshot69 shared benchmarks4–63 head-to-head

Benchmark	Kimi K2.5	Kimi K2.6
AA Agentic Index	52.8	58.7
AA Intelligence	36	45.1
AA-LCR	73	76.7
AA-Omniscience	-7.3	6.4
aime_2026	95.8	96.4
apexAgents	11.5	28.5
arena_elo	1450	1461
arena_text_factuality	1445	1455
arena_vision	1247	1264
Artificial Analysis Coding Index	46.8	61.8
baby_vision	36.5	68.5
baby_vision_with_python	40.5	68.5
browsecomp	74.9	86.3
BrowseComp (Agent Swarm)	78.4	86.3
BrowseComp (w/ Ctx)	74.9	83.2
charxiv_rq	78.7	86.7
charxiv_rq_with_python	78.7	86.7
coding_arena_elo	1436	1509
critpt	3.1	8
deepsearchqa_accuracy	77.1	83
deepsearchqa_f1	89	92.5
FORTRESS (Adversarial)	54.1	65.6
FORTRESS (Benign)	98.3	97.2
frontiermath_tier_4	4.2	14.6
gdpval	38.3	41.4
GDPval-AA v2	1009	1190
Global-MMLU-Lite	84	88.4
GPQA Diamond	87.9	91.1
HLE	50.2	37.5
HLE (with tools)	51.8	54
hmmt_feb_2026	87.1	94.7
IFBench	70.2	76
imo_answer_bench	81.8	86
livebench	69.1	72.2
LiveCodeBench v6	85	89.6
mathvision	84.2	93.2
mathvision_with_python	85	93.2
MCP Atlas	64	68.1
mcpmark	29.5	55.9
MMLU-Pro	87.1	87.1
mmmu_pro_with_python	77.7	80.1
MMMU-Pro	78.5	80.1
OJBench (python)	54.7	60.6
OmniScience Accuracy	35.2	32.8
OmniScience Non-Hallucination	50	60.7
OSWorld-Verified	63.3	73.1
scicode	49	53.5
simpleqa_verified	36.9	38.7
strongreject	99.5	99.8
SWE-bench Multilingual	73	76.7
SWE-bench Pro	53.8	58.6
SWE-bench Verified	76.8	80.2
SWEBench Pro (Public)	50.7	58.6
Tau 3 Banking	14.2	20.6
TauBench V3 - Banking	14.2	23.3
Terminal Bench 2.1 (Best Harness)	51.3	71.3
Terminal-Bench 2.0	50.8	66.7
Terminal-Bench 2.1	45.7	65.9
Terminal-Bench Hard	34.8	43.9
toolathlon	27.8	50
V* (w/ python)	86.9	96.9
vectara_answer_rate	92.2	99.7
vectara_avg_summary_length	112	116.7
vectara_factual_consistency	85.8	89.2
vectara_hallucination_rate ↓	14.2	10.8
WideSearch	79	80.8
WideSearch (item-f1)	72.7	80.8
τ²-Bench Telecom (AA run)	95.9	95.9
τ³-Bench	66	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.