Which is better, GPT-5.6 Sol or Kimi K3?

Across 65 shared benchmarks tracked on Vector Wire, GPT-5.6 Sol leads on 35 and Kimi K3 leads on 30 — GPT-5.6 Sol leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

GPT-5.6 Sol vs Kimi K3

OpenAIvsMoonshot65 shared benchmarks35–30 head-to-head

Benchmark	GPT-5.6 Sol	Kimi K3
AA Agentic Index	57.8	54.3
AA Intelligence	61	60
AA-LCR	77.7	82.7
AA-Omniscience	22	46
Agents' Last Exam	52.7	28.3
aime_2026	99.9	96.7
apex_agents	56.7	41
arc_agi_1	97.5	65.7
ARC-AGI-2	92.5	12.4
arena_elo	1482	1489
arena_text_factuality	1474	1471
Artificial Analysis Coding Index	78.3	76.2
baby_vision_with_python	88.9	85.7
browsecomp	90.4	91.2
charxiv_rq	87.8	84.8
coding_arena_elo	1619	1674
CorpFin v2	64.4	71.6
critpt	32.3	23.4
cybergym	83.6	80
DeepSWE	73	67.5
DeepSWE 1.1	73	69
Finance Agent v2	53.8	54.4
gdpval	61.1	59.4
GDPval-AA v2	1748	1687
GDPval-AA v2 (Elo)	1736	1686
GPQA Diamond	94.6	93.5
HLE	49.5	56
itbenchSre	56.2	47.7
JobBench	45.4	54.3
Legal Research Bench	48.1	44.2
livebench_agentic_coding	56.2	62.2
livebench_coding	83.9	81.5
livebench_data_analysis	79.8	78.7
livebench_instruction_following	71.8	71.4
livebench_language	87.7	85.5
livebench_math	96.2	84.4
livebench_reasoning	91.7	90.7
mathvision	95.8	97.8
MCP Atlas	83.6	84.2
MCP Mark Verified	92.9	94.5
MLS Bench Lite	46.2	48.3
MMMU-Pro	84.6	83.4
MMVU	81.2	82.1
officeqa_pro	63.2	63.3
OmniDocBench	85.8	91.1
OmniScience Accuracy	59.4	47.6
OmniScience Non-Hallucination	10.6	49.1
OSWorld 2.0	62.6	58.3
OSWorld-Verified	83	84.8
PerceptionBench	59.7	58.5
PostTrainBench	34.6	36.6
Program Bench	77.6	77.8
ResearchRubrics	73.8	76.2
SaaS-Bench	61.4	60.1
scicode	56.9	58.7
simplebench	64.8	60.7
simpleqa_verified	71.6	42.7
SpreadsheetBench 2	32.4	34.8
SWE-Marathon	39	42
Tau 3 Banking	33	33.4
TauBench V3 - Banking	44.3	46
Terminal-Bench 2.1	89.5	88.3
toolathlon	58	73.2
Video-MME	89.5	90
WorldVQA ForceAnswer	41.8	51

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.