Which is better, Gemini 3.1 Pro or Gemini 3 Pro?

Across 62 shared benchmarks tracked on Vector Wire, Gemini 3.1 Pro leads on 51 and Gemini 3 Pro leads on 9 — Gemini 3.1 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 3.1 Pro vs Gemini 3 Pro

GooglevsGoogle62 shared benchmarks51–9 head-to-head

Benchmark	Gemini 3.1 Pro	Gemini 3 Pro
AA Agentic Index	23	52
AA Intelligence	47.7	40.6
AA-LCR	79	73
AA-Omniscience	32.9	15.3
aime_2026	98.3	91.7
apexAgents	33.5	18.4
arc_agi_1	98	75
ARC-AGI-2	77.1	54
arena_elo	1486	1485
arena_text_factuality	1471	1481
arena_vision	1280	1290
Artificial Analysis Coding Index	68.8	46.5
browsecomp	85.9	59.2
browsecomp_with_context_manager	85.9	59.2
charxiv_rq	89.9	81.4
coding_arena_elo	1447	1438
critpt	17.7	9.1
cybergym	38.8	39.9
deepsearchqa_f1	81.9	63.2
Fortress	29.8	41.7
frontiermath_tier_4	16.7	18.8
gdpval	23.3	34.2
GDPval-AA (Elo)	1317	1195
GPQA Diamond	94.3	91.9
HLE	51.4	45.8
HLE (with tools)	51.6	45.8
hmmt_feb_2026	94.7	86.4
hmmt_nov_2025	94.8	93.3
IFBench	77.1	70.4
imo_answer_bench	91	83.3
livebench	79.9	73.4
LiveCodeBench	91.7	92
LiveCodeBench Pro (Elo)	2887	2439
LiveCodeBench v6	91.7	90.7
matharena_visual_math_overall	89.4	84.2
mathvision	89.8	86.1
MCP Atlas	78.2	70.3
MMLU-Pro	91	90.1
mmmlu	92.6	91.8
MMMU-Pro	83	81
mrcr_v2_8needle_128k_average	84.9	77
mrcr_v2_8needle_1m_pointwise	26.3	26.3
multichallenge	71.4	65.7
MultiNRC	64.7	59
OmniScience Accuracy	55.3	55.8
OmniScience Non-Hallucination	50.1	10
scicode	59	56.1
simplebench	79.6	76.4
simpleqa_verified	77.3	72.9
SWE-bench Multilingual	76.9	68.7
SWE-bench Pro	54.2	43.3
SWE-bench Verified	80.6	78
Terminal-Bench 2.0	68.5	54.2
Terminal-Bench Hard	68.5	56.9
tool_decathlon	48.8	36.4
toolathlon	48.8	36.4
vectara_answer_rate	99.4	99.4
vectara_avg_summary_length	107.7	101.9
vectara_factual_consistency	89.6	86.4
vectara_hallucination_rate ↓	10.4	13.6
VisualToolBench	29	26.9
τ²-Bench Telecom (AA run)	99.3	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.