Which is better, GPT-5 or Kimi K2 (Reasoning)?

Across 70 shared benchmarks tracked on Vector Wire, GPT-5 leads on 49 and Kimi K2 (Reasoning) leads on 18 — GPT-5 leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

GPT-5 vs Kimi K2 (Reasoning)

OpenAIvsMoonshot70 shared benchmarks49–18 head-to-head

Benchmark	GPT-5	Kimi K2 (Reasoning)
AA Agentic Index	49.7	47.9
AA Intelligence	35.3	33.5
AA-LCR	76.3	70.3
AA-Omniscience	-8.7	-21.4
aider_polyglot	88	59.1
AIME 2025	95	94.5
AIME25 heavy	100	100
AIME25 no tools	94.6	94.5
AIME25 w/ python	99.6	99.1
Arena-Hard (Creative Writing)	92.2	80.1
Arena-Hard (Hard Prompt)	71.9	71.9
ArtifactsBench	73	54.2
Artificial Analysis Coding Index	38.9	34.8
browsecomp	54.9	60.2
BrowseComp w/ tools	54.9	60.2
browsecomp_zh	65	62.3
BrowseComp-ZH w/ tools	63	62.3
critpt	5.7	2.6
FinSearchComp-global	63.9	29.5
FinSearchComp-T3	48.5	47.4
FinSearchComp-T3 w/ tools	48.5	47.4
Frames	86	87
Frames w/ tools	86	87
frontiermath_tier_4	12.5	0
GAIA (text only)	76.4	60.2
gdpval	32.5	24.5
GPQA (unspecified)	85.7	74.2
GPQA Diamond	87.3	84.5
HealthBench	67.2	58
HealthBench no tools	67.2	58
HLE	28.5	23.9
HLE (with tools)	35.2	44.9
HMMT 2025	93.3	89.4
hmmt_feb_2025	88.3	93.3
hmmt_nov_2025	89.2	89.2
HMMT25 heavy	100	97.5
HMMT25 no tools	93.3	89.4
HMMT25 w/ python	96.7	95.1
IFBench	73.1	68.1
imo_answer_bench	76	78.6
LiveCodeBench	85	79.2
LiveCodeBench v6	87	83.1
LiveCodeBenchV6 no tools	87	83.1
Longform Writing eval (Kimi K2 Thinking system card)	71.4	73.8
mmlu_redux	95.3	94.4
MMLU-Pro	87.5	84.6
Multi-SWE-Bench	39.3	41.9
multichallenge	71.1	66.4
OJ-Bench (cpp)	56.2	48.7
OJ-Bench (cpp) no tools	56.2	48.7
OmniScience Accuracy	40.3	30.9
OmniScience Non-Hallucination	21	25.8
scicode	43	44.8
Seal-0	51.4	56.3
Seal-0 w/ tools	51.4	56.3
simplebench	56.7	39.6
swe_bench_bash	65	63.4
SWE-bench Multilingual	55.3	61.1
SWE-bench Verified	74.9	71.3
Terminal-Bench	43.8	47.1
Terminal-Bench 2.0	35.2	35.7
Terminal-Bench Hard	37.9	31.1
Terminal-Bench w/ simulated tools (JSON)	43.8	47.1
vectara_answer_rate	99.9	98.6
vectara_avg_summary_length	162.7	59.2
vectara_factual_consistency	85.3	82.1
vectara_hallucination_rate ↓	14.7	17.9
xbench_deepsearch	77.8	76
τ²-Bench	85	74.3
τ²-Bench Telecom (AA run)	86.5	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.