Which is better, Claude Sonnet 4.6 or Kimi K2.6?

Across 60 shared benchmarks tracked on Vector Wire, Claude Sonnet 4.6 leads on 39 and Kimi K2.6 leads on 21 — Claude Sonnet 4.6 leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

Claude Sonnet 4.6 vs Kimi K2.6

AnthropicvsMoonshot60 shared benchmarks39–21 head-to-head

Benchmark	Claude Sonnet 4.6	Kimi K2.6
AA Agentic Index	61.6	58.7
AA Intelligence	48.4	45.1
AA-LCR	74	76.7
AA-Omniscience	12.2	6.4
AgentWorldBench - Android	58	58.9
AgentWorldBench - MCP	70	65.2
AgentWorldBench - OS	63.2	60.8
AgentWorldBench - Overall	56	53.4
AgentWorldBench - Search	28.8	27.5
AgentWorldBench - SWE	64.5	58.8
AgentWorldBench - Terminal	57	52.5
AgentWorldBench - Web	50.8	50.2
apexAgents	28	28.5
arena_elo	1472	1461
arena_text_factuality	1460	1455
arena_vision	1278	1264
Artificial Analysis Coding Index	63	61.8
browsecomp	76.2	86.3
charxiv_rq	72.4	86.7
coding_arena_elo	1523	1509
critpt	3.1	8
Finance Agent v2	51	44.9
frontiermath_tier_4	8.3	14.6
gdpval	54.8	41.4
GDPval-AA (Elo)	1676	1482
GDPval-AA v2	1381	1190
GPQA Diamond	89.9	91.1
HLE	49	37.5
HLE (with tools)	46.8	54
IFBench	56.6	76
itbenchSre	39.8	31.2
livebench	75.5	72.2
livebench_agentic_coding	42.6	46.9
livebench_coding	79.3	78.6
livebench_data_analysis	78	65.1
livebench_instruction_following	63.2	64.4
livebench_language	76.1	75.1
livebench_math	87	84.3
livebench_reasoning	84.8	79.4
MCP Atlas	69.5	68.1
MMMU-Pro	75.6	80.1
OmniScience Accuracy	40.9	32.8
OmniScience Non-Hallucination	51.6	60.7
OSWorld-Verified	78.5	73.1
scicode	47	53.5
simpleqa_verified	29	38.7
SWE-bench Pro	58.1	58.6
SWE-bench Verified	79.6	80.2
TauBench V3 - Banking	34.4	23.3
Terminal-Bench 2.0	59.1	66.7
Terminal-Bench 2.1	71.2	65.9
Terminal-Bench Hard	59.1	43.9
toolathlon	49.4	50
usamo_2026	55	51.2
vectara_answer_rate	99.9	99.7
vectara_avg_summary_length	114.7	116.7
vectara_factual_consistency	89.4	89.2
vectara_hallucination_rate ↓	10.6	10.8
τ²-Bench Telecom (AA run)	97.9	95.9
τ³-Bench	30.5	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.