Which is better, Claude Opus 4.5 or Kimi K2.6?

Across 53 shared benchmarks tracked on Vector Wire, Claude Opus 4.5 leads on 18 and Kimi K2.6 leads on 35 — 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 —

Claude Opus 4.5 vs Kimi K2.6

AnthropicvsMoonshot53 shared benchmarks18–35 head-to-head

Benchmark	Claude Opus 4.5	Kimi K2.6
AA Agentic Index	59.6	58.7
AA Intelligence	41.9	45.1
AA-LCR	76	76.7
AA-Omniscience	14	6.4
aime_2026	95.1	96.4
arena_elo	1473	1461
arena_text_factuality	1458	1455
Artificial Analysis Coding Index	47.8	61.8
browsecomp	37	86.3
charxiv_rq	68.5	86.7
Claw Eval (pass@3)	59.6	80.9
coding_arena_elo	1495	1509
critpt	4.6	8
deepsearchqa_f1	76.1	92.5
frontiermath_tier_4	4.2	14.6
gdpval	47.3	41.4
GPQA Diamond	87	91.1
HLE	30.8	37.5
HLE (with tools)	43.4	54
hmmt_feb_2026	85.3	94.7
IFBench	58	76
imo_answer_bench	84	86
livebench	76	72.2
livebench_agentic_coding	39.7	46.9
livebench_coding	79.7	78.6
livebench_data_analysis	74.4	65.1
livebench_instruction_following	62.5	64.4
livebench_language	81.3	75.1
livebench_math	90.4	84.3
livebench_reasoning	80.1	79.4
LiveCodeBench	87	89.6
LiveCodeBench v6	84.8	89.6
mathvision	77.1	93.2
MCP Atlas	69.8	68.1
MMLU-Pro	90	87.1
MMMU-Pro	74	80.1
OmniScience Accuracy	46.6	32.8
OmniScience Non-Hallucination	39	60.7
OSWorld-Verified	66.3	73.1
scicode	50	53.5
simpleqa_verified	41.8	38.7
SWE-bench Multilingual	77.5	76.7
SWE-bench Pro	57.1	58.6
SWE-bench Verified	81.5	80.2
Terminal-Bench 2.0	59.3	66.7
Terminal-Bench Hard	47	43.9
toolathlon	43.5	50
vectara_answer_rate	98.7	99.7
vectara_avg_summary_length	114.5	116.7
vectara_factual_consistency	89.1	89.2
vectara_hallucination_rate ↓	10.9	10.8
WideSearch (item-f1)	76.2	80.8
τ²-Bench Telecom (AA run)	89.5	95.9

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.