Which is better, GPT-5.2 or Kimi K2.6?

Across 54 shared benchmarks tracked on Vector Wire, GPT-5.2 leads on 22 and Kimi K2.6 leads on 30 — 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 —

GPT-5.2 vs Kimi K2.6

OpenAIvsMoonshot54 shared benchmarks22–30 head-to-head

Benchmark	GPT-5.2	Kimi K2.6
AA Agentic Index	60.2	58.7
AA Intelligence	43.3	45.1
AA-LCR	79.3	76.7
AA-Omniscience	0.3	6.4
aime_2026	98.3	96.4
apexAgents	23	28.5
arena_elo	1437	1461
arena_text_factuality	1442	1455
arena_vision	1245	1264
Artificial Analysis Coding Index	48.7	61.8
browsecomp	65.8	86.3
charxiv_rq	82.1	86.7
coding_arena_elo	1418	1509
critpt	11.6	8
deepsearchqa_f1	71.3	92.5
frontiermath_tier_4	18.8	14.6
gdpval	48.3	41.4
GDPval-AA (Elo)	1462	1482
GPQA Diamond	92.4	91.1
HLE	45.5	37.5
HLE (with tools)	45.5	54
hmmt_feb_2026	97	94.7
IFBench	75.4	76
imo_answer_bench	86.3	86
livebench	74.8	72.2
livebench_agentic_coding	50.3	46.9
livebench_coding	76.1	78.6
livebench_data_analysis	78.2	65.1
livebench_instruction_following	61.8	64.4
livebench_language	79.8	75.1
livebench_math	93.2	84.3
livebench_reasoning	83.2	79.4
LiveCodeBench	89	89.6
mathvision	83	93.2
MCP Atlas	68	68.1
MMLU-Pro	87	87.1
MMMU-Pro	79.5	80.1
OmniScience Accuracy	44.3	32.8
OmniScience Non-Hallucination	38.4	60.7
scicode	52.1	53.5
simpleqa_verified	38.9	38.7
SWE-bench Multilingual	72	76.7
SWE-bench Pro	55.6	58.6
SWE-bench Verified	80	80.2
TauBench V3 - Banking	11.1	23.3
Terminal-Bench 2.0	54	66.7
Terminal-Bench Hard	62.2	43.9
toolathlon	46.3	50
vectara_answer_rate	100	99.7
vectara_avg_summary_length	186.3	116.7
vectara_factual_consistency	89.2	89.2
vectara_hallucination_rate ↓	10.8	10.8
τ²-Bench Telecom (AA run)	98.7	95.9
τ³-Bench	12	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.