Which is better, Gemini 3.1 Pro or GPT-5.4?

Across 94 shared benchmarks tracked on Vector Wire, Gemini 3.1 Pro leads on 41 and GPT-5.4 leads on 53 — 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 —

Gemini 3.1 Pro vs GPT-5.4

GooglevsOpenAI94 shared benchmarks41–53 head-to-head

Benchmark	Gemini 3.1 Pro	GPT-5.4
AA Agentic Index	23	58.2
AA Intelligence	47.7	53.1
AA-LCR	79	77.7
AA-Omniscience	32.9	5.8
AgentWorldBench - Android	61.4	60
AgentWorldBench - MCP	59.1	70.1
AgentWorldBench - OS	66.9	68.6
AgentWorldBench - Overall	54.6	58.3
AgentWorldBench - Search	30.2	37.3
AgentWorldBench - SWE	59.1	66.3
AgentWorldBench - Terminal	52.5	53.7
AgentWorldBench - Web	52.8	51.8
aime_2026	98.3	99.2
Apex (Pass@1)	60.9	54.1
Apex Shortlist (Pass@1)	89.1	78.1
apexAgents	33.5	33.3
arc_agi_1	98	93.7
arc_agi_3	0.4	0.2
ARC-AGI-2	77.1	74
arena_elo	1486	1477
arena_text_factuality	1471	1476
arena_vision	1280	1277
Artificial Analysis Coding Index	68.8	71.1
baby_vision	51.6	49.7
baby_vision_with_python	68.3	80.2
browsecomp	85.9	82.7
BrowseComp (Agent Swarm)	85.9	82.7
browsecomp_with_context_manager	85.9	82.7
charxiv_rq	89.9	82.8
charxiv_rq_with_python	89.9	90
Chinese SimpleQA (C-SimpleQA)	85.9	76.8
Codeforces	3052	3168
coding_arena_elo	1447	1457
critpt	17.7	23.4
cybergym	38.8	66.3
deepsearchqa_accuracy	60.2	63.7
deepsearchqa_f1	81.9	78.6
DeepSWE 1.1	12	52
finance_agent	59.7	57.2
frontiermath_tier_4	16.7	27.1
gdpval	23.3	50.1
GDPval-AA (Elo)	1317	1674
GPQA Diamond	94.3	93
HiL-Bench	35.3	9.7
HLE	51.4	43.7
HLE (with tools)	51.6	52.1
hmmt_feb_2026	94.7	97.7
hmmt_nov_2025	94.8	95.8
IFBench	77.1	73.9
imo_answer_bench	91	91.4
itbenchSre	30.3	34.5
Legal Agent Benchmark	0	0.4
livebench	79.9	80.3
livebench_agentic_coding	44.1	53.8
livebench_coding	76.5	77.5
livebench_data_analysis	78.5	79.3
livebench_instruction_following	79.1	70.2
livebench_language	85.4	82.6
livebench_math	91	94.2
livebench_reasoning	84	88.1
matharena_visual_math_overall	89.4	92.5
mathvision	89.8	92
mathvision_with_python	95.7	96.1
MCP Atlas	78.2	70.6
MCPAtlas Public (Pass@1)	69.2	67.2
mcpmark	55.9	62.5
MMLU-Pro	91	87.5
mmmu_pro_with_python	85.3	82.1
MMMU-Pro	83	81.2
MultiNRC	64.7	58.3
nl2repo	33.4	41.3
officeqa_pro	42.9	51.1
OmniScience Accuracy	55.3	50.9
OmniScience Non-Hallucination	50.1	17.4
OSWorld-Verified	76.2	75
scicode	59	56.6
simpleqa_verified	77.3	45.3
SWE-bench Multilingual	76.9	71.7
SWE-bench Pro	54.2	59.1
TauBench V3 - Banking	21.4	39.6
Terminal-Bench 2.0	68.5	75.1
Terminal-Bench 2.1	74	78.3
Terminal-Bench Hard	68.5	57.6
tool_decathlon	48.8	54.6
toolathlon	48.8	54.6
usamo_2026	74.4	95.2
V* (w/ python)	96.9	98.4
vectara_answer_rate	99.4	99.9
vectara_avg_summary_length	107.7	81.7
vectara_factual_consistency	89.6	93
vectara_hallucination_rate ↓	10.4	7
VisualToolBench	29	29.2
τ²-Bench Telecom (AA run)	99.3	87.1
τ³-Bench	67.1	72.9

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.