Which is better, DeepSeek-R1 or GPT-4o?

Across 63 shared benchmarks tracked on Vector Wire, DeepSeek-R1 leads on 50 and GPT-4o leads on 13 — DeepSeek-R1 leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

DeepSeek-R1 vs GPT-4o

DeepSeekvsOpenAI63 shared benchmarks50–13 head-to-head

Benchmark	DeepSeek-R1	GPT-4o
AA Agentic Index	3.1	9.7
AA Intelligence	18.6	11.2
AA-LCR	56	55
AA-Omniscience	-31.3	-10.5
aider_polyglot	71.6	30.7
AIME 2024	91.4	9.3
AIME 2025	87.5	11.7
AIR-Bench 2024	52.9	62.4
AlpacaEval2.0 (LC-winrate)	87.6	51.1
anthropic_red_team	99.1	99.1
arc_agi_1	21.2	4.5
ARC-AGI-2	1.3	0
Arena-Hard (GPT-4-1106 judge)	92.3	80.4
ArenaHard (GPT-4-1106)	92.3	80.4
Artificial Analysis Coding Index	24.6	24.2
bbq	96.6	95.1
C-Eval	91.8	76
Chinese SimpleQA (C-SimpleQA)	63.7	64.6
CLUEWSC	92.8	87.9
CNMO 2024	78.8	10.8
Codeforces	2029	759
Codeforces (Percentile)	96.3	23.6
Codeforces (Rating)	2029	759
critpt	0.6	0
DROP (3-shot F1)	92.2	83.7
Fortress	74.4	47.2
FRAMES (Acc.)	83	80.5
gdpval	1.5	0
GPQA Diamond	81	70.1
harmbench	54.6	82.9
HLE	17.7	5.3
IFBench	39	36
ifeval	83.3	84.3
LiveCodeBench	84.4	38.3
livecodebench_easy	99.2	82.5
livecodebench_hard	63.6	4.5
livecodebench_medium	90.9	32.1
livecodebench_pass1cot	65.9	34.2
longbench_v2	58.3	48.1
MATH	97.3	85.3
math_500_em	98	74.6
mmlu	92.9	88.1
mmlu_prox	75.5	61.1
mmlu_redux	93.4	88
MMLU-Pro	85	74.7
multichallenge	45	40.3
MultiNRC	27.6	12.4
OmniScience Accuracy	30.7	23.7
OmniScience Non-Hallucination	10.5	62.1
scicode	35.7	33.4
simple_safety_tests	98.3	98.5
simplebench	40.8	17.8
simpleqa	92.3	39.4
SWE-bench Verified	57.6	38.8
TAU-bench (airline)	53.5	42.8
TAU-bench (retail)	63.9	60.3
Terminal-Bench Hard	6.1	8.3
vectara_answer_rate	97	93.8
vectara_avg_summary_length	93.5	86.6
vectara_factual_consistency	88.7	90.4
vectara_hallucination_rate ↓	11.3	9.6
xstest	98.8	97.3
τ²-Bench Telecom (AA run)	11.4	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.