Which is better, GPT-5.4 or Kimi K2.6?

Across 79 shared benchmarks tracked on Vector Wire, GPT-5.4 leads on 61 and Kimi K2.6 leads on 18 — 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 K2.6

OpenAIvsMoonshot79 shared benchmarks61–18 head-to-head

Benchmark	GPT-5.4	Kimi K2.6
AA Agentic Index	58.2	58.7
AA Intelligence	53.1	45.1
AA-LCR	77.7	76.7
AA-Omniscience	5.8	6.4
AgentWorldBench - Android	60	58.9
AgentWorldBench - MCP	70.1	65.2
AgentWorldBench - OS	68.6	60.8
AgentWorldBench - Overall	58.3	53.4
AgentWorldBench - Search	37.3	27.5
AgentWorldBench - SWE	66.3	58.8
AgentWorldBench - Terminal	53.7	52.5
AgentWorldBench - Web	51.8	50.2
aime_2026	99.2	96.4
Apex (Pass@1)	54.1	24
Apex Shortlist (Pass@1)	78.1	75.5
apexAgents	33.3	28.5
arena_elo	1477	1461
arena_text_factuality	1476	1455
arena_vision	1277	1264
Artificial Analysis Coding Index	71.1	61.8
baby_vision	49.7	68.5
baby_vision_with_python	80.2	68.5
browsecomp	82.7	86.3
BrowseComp (Agent Swarm)	82.7	86.3
charxiv_rq	82.8	86.7
charxiv_rq_with_python	90	86.7
Chinese SimpleQA (C-SimpleQA)	76.8	75.9
coding_arena_elo	1457	1509
critpt	23.4	8
deepsearchqa_accuracy	63.7	83
deepsearchqa_f1	78.6	92.5
frontiermath_tier_4	27.1	14.6
gdpval	50.1	41.4
GDPval-AA (Elo)	1674	1482
GPQA Diamond	93	91.1
HiL-Bench	9.7	18.7
HLE	43.7	37.5
HLE (with tools)	52.1	54
hmmt_feb_2026	97.7	94.7
IFBench	73.9	76
imo_answer_bench	91.4	86
itbenchSre	34.5	31.2
livebench	80.3	72.2
livebench_agentic_coding	53.8	46.9
livebench_coding	77.5	78.6
livebench_data_analysis	79.3	65.1
livebench_instruction_following	70.2	64.4
livebench_language	82.6	75.1
livebench_math	94.2	84.3
livebench_reasoning	88.1	79.4
mathvision	92	93.2
mathvision_with_python	96.1	93.2
MCP Atlas	70.6	68.1
MCPAtlas Public (Pass@1)	67.2	66.6
mcpmark	62.5	55.9
MMLU-Pro	87.5	87.1
mmmu_pro_with_python	82.1	80.1
MMMU-Pro	81.2	80.1
OmniScience Accuracy	50.9	32.8
OmniScience Non-Hallucination	17.4	60.7
OSWorld-Verified	75	73.1
scicode	56.6	53.5
simpleqa_verified	45.3	38.7
SWE-bench Multilingual	71.7	76.7
SWE-bench Pro	59.1	58.6
SWE-QA	81.3	71.6
TauBench V3 - Banking	39.6	23.3
Terminal-Bench 2.0	75.1	66.7
Terminal-Bench 2.1	78.3	65.9
Terminal-Bench Hard	57.6	43.9
toolathlon	54.6	50
usamo_2026	95.2	51.2
V* (w/ python)	98.4	96.9
vectara_answer_rate	99.9	99.7
vectara_avg_summary_length	81.7	116.7
vectara_factual_consistency	93	89.2
vectara_hallucination_rate ↓	7	10.8
τ²-Bench Telecom (AA run)	87.1	95.9
τ³-Bench	72.9	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.