Which is better, Gemini 3 Pro or GPT-5.4?

Across 54 shared benchmarks tracked on Vector Wire, Gemini 3 Pro leads on 12 and GPT-5.4 leads on 42 — 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 Pro vs GPT-5.4

GooglevsOpenAI54 shared benchmarks12–42 head-to-head

Benchmark	Gemini 3 Pro	GPT-5.4
AA Agentic Index	52	58.2
AA Intelligence	40.6	53.1
AA-LCR	73	77.7
AA-Omniscience	15.3	5.8
aime_2026	91.7	99.2
apexAgents	18.4	33.3
arc_agi_1	75	93.7
ARC-AGI-2	54	74
arena_elo	1485	1477
arena_text_factuality	1481	1476
arena_vision	1290	1277
Artificial Analysis Coding Index	46.5	71.1
browsecomp	59.2	82.7
browsecomp_with_context_manager	59.2	82.7
charxiv_rq	81.4	82.8
coding_arena_elo	1438	1457
critpt	9.1	23.4
cybergym	39.9	66.3
deepsearchqa_f1	63.2	78.6
frontiermath_tier_4	18.8	27.1
gdpval	34.2	50.1
GDPval-AA (Elo)	1195	1674
GPQA Diamond	91.9	93
HLE	45.8	43.7
HLE (with tools)	45.8	52.1
hmmt_feb_2026	86.4	97.7
hmmt_nov_2025	93.3	95.8
IFBench	70.4	73.9
imo_answer_bench	83.3	91.4
livebench	73.4	80.3
matharena_visual_math_overall	84.2	92.5
mathvision	86.1	92
MCP Atlas	70.3	70.6
MMLU-Pro	90.1	87.5
MMMU-Pro	81	81.2
MultiNRC	59	58.3
OmniDocBench 1.5	88.5	89.1
OmniScience Accuracy	55.8	50.9
OmniScience Non-Hallucination	10	17.4
scicode	56.1	56.6
SEAL VISTA	51.5	50.9
simpleqa_verified	72.9	45.3
SWE-bench Multilingual	68.7	71.7
SWE-bench Pro	43.3	59.1
Terminal-Bench 2.0	54.2	75.1
Terminal-Bench Hard	56.9	57.6
tool_decathlon	36.4	54.6
toolathlon	36.4	54.6
vectara_answer_rate	99.4	99.9
vectara_avg_summary_length	101.9	81.7
vectara_factual_consistency	86.4	93
vectara_hallucination_rate ↓	13.6	7
VisualToolBench	26.9	29.2
τ²-Bench Telecom (AA run)	98	87.1

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.