Which is better, Claude Sonnet 4.5 or Kimi K2 (Reasoning)?

Across 70 shared benchmarks tracked on Vector Wire, Claude Sonnet 4.5 leads on 33 and Kimi K2 (Reasoning) leads on 37 — Kimi K2 (Reasoning) 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 (Reasoning)

AnthropicvsMoonshot70 shared benchmarks33–37 head-to-head

Benchmark	Claude Sonnet 4.5	Kimi K2 (Reasoning)
AA Agentic Index	50.6	47.9
AA Intelligence	37.4	33.5
AA-LCR	68.3	70.3
AA-Omniscience	-0.1	-21.4
AIME 2025	88	94.5
AIME25 no tools	88	94.5
AIME25 w/ python	100	99.1
Arena-Hard (Creative Writing)	76.7	80.1
Arena-Hard (Hard Prompt)	63.3	71.9
ArtifactsBench	61.5	54.2
Artificial Analysis Coding Index	52.1	34.8
browsecomp	24.1	60.2
BrowseComp w/ tools	24.1	60.2
browsecomp_zh	42.4	62.3
BrowseComp-ZH w/ tools	42.4	62.3
critpt	1.1	2.6
FinSearchComp-global	60.8	29.5
FinSearchComp-T3	44	47.4
FinSearchComp-T3 w/ tools	44	47.4
Frames	85	87
Frames w/ tools	85	87
frontiermath_tier_4	4.2	0
GAIA (text only)	71.2	60.2
gdpval	40.4	24.5
GPQA (unspecified)	83.4	74.2
GPQA Diamond	83.4	84.5
HealthBench	44.2	58
HealthBench no tools	44.2	58
HLE	19.8	23.9
HLE (with tools)	32	44.9
HMMT 2025	74.6	89.4
hmmt_feb_2025	79.2	93.3
hmmt_nov_2025	81.7	89.2
HMMT25 no tools	74.6	89.4
HMMT25 w/ python	88.8	95.1
IFBench	57.3	68.1
imo_answer_bench	65.9	78.6
LiveCodeBench	71	79.2
LiveCodeBench v6	64	83.1
LiveCodeBenchV6 no tools	64	83.1
longbench_v2	61.8	45.1
Longform Writing eval (Kimi K2 Thinking system card)	79.8	73.8
mmlu_redux	95.6	94.4
MMLU-Pro	88.2	84.6
mrcr	55.4	44.2
Multi-SWE-Bench	44.3	41.9
multichallenge	55.3	66.4
OJ-Bench (cpp)	30.4	48.7
OJ-Bench (cpp) no tools	30.4	48.7
OmniScience Accuracy	32.9	30.9
OmniScience Non-Hallucination	50.9	25.8
scicode	45	44.8
Seal-0	53.4	56.3
Seal-0 w/ tools	53.4	56.3
simplebench	54.3	39.6
swe_bench_bash	71.4	63.4
SWE-bench Multilingual	68	61.1
SWE-bench Verified	82	71.3
Terminal-Bench	51	47.1
Terminal-Bench 2.0	50	35.7
Terminal-Bench Hard	35.6	31.1
Terminal-Bench w/ simulated tools (JSON)	51	47.1
theagentcompany	41	30
vectara_answer_rate	95.6	98.6
vectara_avg_summary_length	127.8	59.2
vectara_factual_consistency	88	82.1
vectara_hallucination_rate ↓	12	17.9
xbench_deepsearch	66	76
τ²-Bench	98	74.3
τ²-Bench Telecom (AA run)	78.1	93

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.