Which is better, Claude Sonnet 4.5 or Kimi K2.5?

Across 64 shared benchmarks tracked on Vector Wire, Claude Sonnet 4.5 leads on 24 and Kimi K2.5 leads on 40 — Kimi K2.5 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 4.5 vs Kimi K2.5

AnthropicvsMoonshot64 shared benchmarks24–40 head-to-head

Benchmark	Claude Sonnet 4.5	Kimi K2.5
AA Agentic Index	50.6	52.8
AA Intelligence	37.4	36
AA-LCR	68.3	73
AA-Omniscience	-0.1	-7.3
AIME 2025	88	96.1
arc_agi_1	63.7	65.3
ARC-AGI-2	13.6	11.8
arena_elo	1456	1450
arena_text_factuality	1448	1445
Artificial Analysis Coding Index	52.1	46.8
browsecomp	24.1	74.9
BrowseComp (context management)	26.1	74.9
browsecomp_zh	42.4	62.3
charxiv_rq	67.2	78.7
coding_arena_elo	1392	1436
critpt	1.1	3.1
Fortress	17	41.1
frontiermath_tier_4	4.2	4.2
gdpval	40.4	38.3
GPQA Diamond	83.4	87.9
HLE	19.8	50.2
HLE (with tools)	32	51.8
HMMT 2025	74.6	95.4
hmmt_feb_2025	79.2	95.4
hmmt_nov_2025	81.7	91.1
IFBench	57.3	70.2
imo_answer_bench	65.9	81.8
LiveCodeBench v6	64	85
longbench_v2	61.8	61
matharena_visual_math_overall	75.8	80.6
MathVista	79.8	90.1
MCP Atlas	59.5	64
MMLU-Pro	88.2	87.1
MMMU-Pro	68.7	78.5
multichallenge	55.3	61.4
MultiNRC	35.8	35.2
OJ-Bench (cpp)	30.4	57.4
OmniDocBench 1.5	85.8	88.8
OmniScience Accuracy	32.9	35.2
OmniScience Non-Hallucination	50.9	50
OSWorld-Verified	61.4	63.3
scicode	45	49
SEAL VISTA	48.8	41.9
Seal-0	53.4	57.4
simplebench	54.3	46.8
simpleqa_verified	23.6	36.9
SimpleVQA	57.6	71.2
swe_bench_bash	71.4	70.8
SWE-bench Multilingual	68	73
SWE-bench Pro	43.6	53.8
SWE-bench Verified	82	76.8
TauBench V3 - Banking	24.5	14.2
Terminal-Bench 2.0	50	50.8
Terminal-Bench 2.1	55.8	45.7
Terminal-Bench Hard	35.6	34.8
toolathlon	41	27.8
vectara_answer_rate	95.6	92.2
vectara_avg_summary_length	127.8	112
vectara_factual_consistency	88	85.8
vectara_hallucination_rate ↓	12	14.2
xbench_deepsearch	66	76.7
τ²-Bench	98	85.4
τ²-Bench Telecom (AA run)	78.1	95.9
τ³-Bench	19	66

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.