Which is better, DeepSeek-V3 or GPT-4o?

Across 76 shared benchmarks tracked on Vector Wire, DeepSeek-V3 leads on 50 and GPT-4o leads on 24 — DeepSeek-V3 leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

DeepSeek-V3 vs GPT-4o

DeepSeekvsOpenAI76 shared benchmarks50–24 head-to-head

Benchmark	DeepSeek-V3	GPT-4o
AA Agentic Index	1.6	9.7
AA Intelligence	15.4	11.2
AA-LCR	41.3	55
AA-Omniscience	-37.6	-10.5
aider_edit_acc	79.7	72.9
aider_polyglot	55.1	30.7
AIME 2024	59.4	9.3
AIME 2025	51.3	11.7
AIR-Bench 2024	40.8	62.4
AlpacaEval2.0 (LC-winrate)	70	51.1
anthropic_red_team	97.1	99.1
Arena Hard	91.4	92.4
Arena-Hard (GPT-4-1106 judge)	85.5	80.4
ArenaHard (GPT-4-1106)	85.5	80.4
Artificial Analysis Coding Index	23	24.2
bbq	96.7	95.1
C-Eval	90.1	76
Chinese SimpleQA (C-SimpleQA)	68	64.6
CLUEWSC	90.9	87.9
CNMO 2024	74.7	10.8
Codeforces	1134	759
Codeforces (Percentile)	58.7	23.6
Codeforces (Rating)	1134	759
critpt	0	0
DROP	91.6	83.4
DROP (3-shot F1)	91.6	83.7
DROP (F1)	91	89.2
FRAMES (Acc.)	73.3	80.5
gdpval	0	0
GPQA Diamond	68.4	70.1
GSM8K	96.7	95.6
harmbench	49.7	82.9
HLE	5.2	5.3
humaneval	92.1	90.6
HumanEval-Mul (Pass@1)	82.6	80.5
IFBench	41	36
ifeval	87.3	84.3
IFEval (avg)	87.3	84.1
LiveCodeBench	49.6	38.3
LiveCodeBench v6	46.9	30.9
livecodebench_easy	83.3	82.5
livecodebench_hard	14.2	4.5
livecodebench_medium	53.5	32.1
livecodebench_pass1cot	40.5	34.2
LongBench v2 overall (w/o CoT)	48.7	50.1
longbench_v2	48.7	48.1
MATH	91.2	85.3
math_500_em	94	74.6
MBPP+ (EvalPlus-augmented)	78.8	76.2
mgsm	79.8	90.5
mmlu	89.4	88.1
mmlu_prox	70.5	61.1
mmlu_redux	90.5	88
MMLU-Pro	81.2	74.7
mmmlu	79.4	81.4
multichallenge	31.4	40.3
narrativeqa	79.6	80.4
naturalquestions_closedbook	46.7	50.1
OmniScience Accuracy	25.4	23.7
OmniScience Non-Hallucination	23.3	62.1
OpenBookQA	95.4	96.8
scicode	35.8	33.4
simple_safety_tests	95.3	98.5
simplebench	27.2	17.8
simpleqa	27.7	39.4
supergpqa	53.7	42.4
SWE-bench Verified	42	38.8
Tau2 airline	39	45.5
Tau2 retail	69.1	63.4
Terminal-Bench Hard	15.2	8.3
vectara_answer_rate	97.5	93.8
vectara_avg_summary_length	81.7	86.6
vectara_factual_consistency	93.9	90.4
vectara_hallucination_rate ↓	6.1	9.6
xstest	97.1	97.3
τ²-Bench Telecom (AA run)	47.1	28.9

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.