Which is better, Gemini 3 Flash Preview or Gemini 3 Pro?

Across 54 shared benchmarks tracked on Vector Wire, Gemini 3 Flash Preview leads on 16 and Gemini 3 Pro leads on 32 — Gemini 3 Pro 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 Flash Preview vs Gemini 3 Pro

GooglevsGoogle54 shared benchmarks16–32 head-to-head

Benchmark	Gemini 3 Flash Preview	Gemini 3 Pro
AA Agentic Index	49.7	52
AA Intelligence	38.7	40.6
AA-LCR	73	73
AA-Omniscience	10.1	15.3
AIME 2025	99.7	100
aime_2026	96.7	91.7
apexAgents	27.7	18.4
arc_agi_1	84.7	75
ARC-AGI-2	33.6	54
arena_elo	1473	1485
arena_text_factuality	1452	1481
arena_vision	1260	1290
Artificial Analysis Coding Index	42.6	46.5
charxiv_rq	80.3	81.4
coding_arena_elo	1387	1438
critpt	8.6	9.1
frontiermath_tier_4	4.2	18.8
gdpval	35.2	34.2
GDPval-AA (Elo)	1204	1195
GPQA Diamond	90.4	91.9
HLE	43.5	45.8
hmmt_feb_2025	97.5	97.5
hmmt_feb_2026	89.4	86.4
hmmt_nov_2025	93.3	93.3
IFBench	78	70.4
livebench	72.4	73.4
matharena_visual_math_overall	85.8	84.2
MCP Atlas	62	70.3
mmmlu	91.8	91.8
MMMU-Pro	81.2	81
mrcr_v2_8needle_128k_average	67.2	77
mrcr_v2_8needle_1m_pointwise	26.3	26.3
OmniDocBench 1.5	12.1	88.5
OmniScience Accuracy	53.4	55.8
OmniScience Non-Hallucination	7.6	10
scicode	50.6	56.1
screenspot_pro_no_tools	69.1	72.7
simplebench	61.1	76.4
simpleqa	68.7	72.1
simpleqa_verified	67.4	72.9
swe_bench_bash	75.8	74.2
SWE-bench Multilingual	72.7	68.7
SWE-bench Pro	49.6	43.3
SWE-bench Verified	78	78
Terminal-Bench 2.0	47.6	54.2
Terminal-Bench Hard	38.6	56.9
toolathlon	49.4	36.4
vectara_answer_rate	99.8	99.4
vectara_avg_summary_length	90.2	101.9
vectara_factual_consistency	86.5	86.4
vectara_hallucination_rate ↓	13.5	13.6
Vending-Bench 2	363500	547816
VideoMMMU	86.9	87.6
τ²-Bench Telecom (AA run)	80.4	98

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.