Which is better, DeepSeek-V4-Flash or Gemini 3.1 Pro?

Across 60 shared benchmarks tracked on Vector Wire, DeepSeek-V4-Flash leads on 21 and Gemini 3.1 Pro leads on 38 — 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 —

DeepSeek-V4-Flash vs Gemini 3.1 Pro

DeepSeekvsGoogle60 shared benchmarks21–38 head-to-head

Benchmark	DeepSeek-V4-Flash	Gemini 3.1 Pro
AA Agentic Index	61.3	23
AA Intelligence	51.8	47.7
AA-LCR	74.3	79
AA-Omniscience	-14.3	32.9
aime_2026	95.8	98.3
Apex (Pass@1)	33	60.9
Apex Shortlist (Pass@1)	85.7	89.1
arc_agi_1	84	98
ARC-AGI-2	46	77.1
arena_elo	1439	1486
arena_text_factuality	1434	1471
Artificial Analysis Coding Index	69.1	68.8
browsecomp	73.2	85.9
Chinese SimpleQA (C-SimpleQA)	78.9	85.9
Codeforces	3052	3052
coding_arena_elo	1431	1447
CorpusQA 1M (ACC)	60.5	53.8
critpt	16.6	17.7
cybergym	76.7	38.8
DeepSWE	54.4	10
gdpval	52.9	23.3
GDPval-AA (Elo)	1395	1317
GDPval-AA v2	1189	962
Global-MMLU-Lite	88.4	92.7
GPQA Diamond	90.8	94.3
HLE	38.6	51.4
HLE (with tools)	45.1	51.6
hmmt_feb_2026	94.8	94.7
IFBench	79.2	77.1
imo_answer_bench	88.4	91
itbenchSre	31.5	30.3
livebench_agentic_coding	46.8	44.1
livebench_coding	75	76.5
livebench_data_analysis	79.3	78.5
livebench_instruction_following	65.5	79.1
livebench_language	79.2	85.4
livebench_math	86.8	91
livebench_reasoning	86.6	84
LiveCodeBench	91.6	91.7
MCP Atlas	69	78.2
MMLU-Pro	86.4	91
MRCR 1M (MMR)	78.7	76.3
nl2repo	54.2	33.4
OmniScience Accuracy	40.4	55.3
OmniScience Non-Hallucination	15.6	50.1
scicode	49.9	59
simplebench	46.3	79.6
simpleqa_verified	45	77.3
strongreject	97.4	98
SWE-bench Multilingual	73.3	76.9
SWE-bench Pro	52.6	54.2
SWE-bench Verified	79	80.6
Tau 3 Banking	22.9	16.5
TauBench V3 - Banking	39.4	21.4
Terminal-Bench 2.0	56.9	68.5
Terminal-Bench 2.1	82.7	74
Terminal-Bench Hard	38.6	68.5
toolathlon	70.3	48.8
τ²-Bench Telecom (AA run)	95.6	99.3
τ³-Bench	22.9	67.1

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.