Which is better, GPT-5.4 or Kimi K2.5?

Across 66 shared benchmarks tracked on Vector Wire, GPT-5.4 leads on 59 and Kimi K2.5 leads on 7 — GPT-5.4 leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

GPT-5.4 vs Kimi K2.5

OpenAIvsMoonshot66 shared benchmarks59–7 head-to-head

Benchmark	GPT-5.4	Kimi K2.5
AA Agentic Index	58.2	52.8
AA Intelligence	53.1	36
AA-LCR	77.7	73
AA-Omniscience	5.8	-7.3
aime_2026	99.2	95.8
apexAgents	33.3	11.5
arc_agi_1	93.7	65.3
ARC-AGI-2	74	11.8
arena_elo	1477	1450
arena_text_factuality	1476	1445
arena_vision	1277	1247
Artificial Analysis Coding Index	71.1	46.8
baby_vision	49.7	36.5
baby_vision_with_python	80.2	40.5
browsecomp	82.7	74.9
BrowseComp (Agent Swarm)	82.7	78.4
browsecomp_with_context_manager	82.7	74.9
charxiv_rq	82.8	78.7
charxiv_rq_with_python	90	78.7
coding_arena_elo	1457	1436
critpt	23.4	3.1
cybergym	66.3	41.3
deepsearchqa_accuracy	63.7	77.1
deepsearchqa_f1	78.6	89
frontiermath_tier_4	27.1	4.2
gdpval	50.1	38.3
GPQA Diamond	93	87.9
HLE	43.7	50.2
HLE (with tools)	52.1	51.8
hmmt_feb_2026	97.7	87.1
hmmt_nov_2025	95.8	91.1
IFBench	73.9	70.2
imo_answer_bench	91.4	81.8
livebench	80.3	69.1
matharena_visual_math_overall	92.5	80.6
mathvision	92	84.2
mathvision_with_python	96.1	85
MCP Atlas	70.6	64
mcpmark	62.5	29.5
MMLU-Pro	87.5	87.1
mmmu_pro_with_python	82.1	77.7
MMMU-Pro	81.2	78.5
MultiNRC	58.3	35.2
nl2repo	41.3	32
OmniDocBench 1.5	89.1	88.8
OmniScience Accuracy	50.9	35.2
OmniScience Non-Hallucination	17.4	50
OSWorld-Verified	75	63.3
scicode	56.6	49
SEAL VISTA	50.9	41.9
simpleqa_verified	45.3	36.9
SWE-bench Multilingual	71.7	73
SWE-bench Pro	59.1	53.8
TauBench V3 - Banking	39.6	14.2
Terminal-Bench 2.0	75.1	50.8
Terminal-Bench 2.1	78.3	45.7
Terminal-Bench Hard	57.6	34.8
tool_decathlon	54.6	27.8
toolathlon	54.6	27.8
V* (w/ python)	98.4	86.9
vectara_answer_rate	99.9	92.2
vectara_avg_summary_length	81.7	112
vectara_factual_consistency	93	85.8
vectara_hallucination_rate ↓	7	14.2
τ²-Bench Telecom (AA run)	87.1	95.9
τ³-Bench	72.9	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.