Which is better, Gemini 3.1 Pro or GPT-5.6 Sol?

Across 63 shared benchmarks tracked on Vector Wire, Gemini 3.1 Pro leads on 16 and GPT-5.6 Sol leads on 47 — 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 —

Gemini 3.1 Pro vs GPT-5.6 Sol

GooglevsOpenAI63 shared benchmarks16–47 head-to-head

Benchmark	Gemini 3.1 Pro	GPT-5.6 Sol
AA Agentic Index	23	57.8
AA Intelligence	47.7	61
AA-LCR	79	77.7
AA-Omniscience	32.9	22
aime_2026	98.3	99.9
apex_agents	33.5	56.7
arc_agi_1	98	97.5
arc_agi_3	0.4	7.8
ARC-AGI-2	77.1	92.5
arena_elo	1486	1482
arena_text_factuality	1471	1474
Artificial Analysis Coding Index	68.8	78.3
baby_vision_with_python	68.3	88.9
browsecomp	85.9	90.4
BrowseComp (w/ Ctx)	85.9	89.4
charxiv_rq	89.9	87.8
coding_arena_elo	1447	1619
critpt	17.7	32.3
cybergym	38.8	83.6
DeepSWE	10	73
DeepSWE 1.1	12	73
Finance Agent v2	43	53.8
FORTRESS (Adversarial)	65.2	82.4
FORTRESS (Benign)	98	98.1
gdpval	23.3	61.1
GDPval-AA v2	962	1748
Global-MMLU-Lite	92.7	91.8
GPQA Diamond	94.3	94.6
HiL-Bench	35.3	32.3
HLE	51.4	49.5
HLE (with tools)	51.6	58
IFBench	77.1	72.7
itbenchSre	30.3	56.2
livebench_agentic_coding	44.1	56.2
livebench_coding	76.5	83.9
livebench_data_analysis	78.5	79.8
livebench_instruction_following	79.1	71.8
livebench_language	85.4	87.7
livebench_math	91	96.2
livebench_reasoning	84	91.7
mathvision	89.8	95.8
MCP Atlas	78.2	83.6
MMMU-Pro	83	84.6
officeqa_pro	42.9	63.2
OmniScience Accuracy	55.3	59.4
OmniScience Non-Hallucination	50.1	10.6
OSWorld-Verified	76.2	83
PostTrainBench	21.6	34.6
Program Bench	39.5	77.6
scicode	59	56.9
simplebench	79.6	64.8
simpleqa_verified	77.3	71.6
strongreject	98	98.5
SWE-bench Pro	54.2	64.6
SWE-Marathon	4	39
SWEBench Pro (Public)	54.2	64.6
Tau 3 Banking	16.5	33
TauBench V3 - Banking	21.4	44.3
Terminal Bench 2.1 (Best Harness)	73.8	89.5
Terminal-Bench 2.1	74	89.5
Terminal-Bench Hard	68.5	65.9
toolathlon	48.8	58
τ²-Bench Telecom (AA run)	99.3	85.1

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.