Which is better, GPT-4o or o4-mini?

Across 49 shared benchmarks tracked on Vector Wire, GPT-4o leads on 9 and o4-mini leads on 40 — o4-mini leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

GPT-4o vs o4-mini

OpenAIvsOpenAI49 shared benchmarks9–40 head-to-head

Benchmark	GPT-4o	o4-mini
AA Agentic Index	9.7	36.1
AA Intelligence	11.2	26.1
AA-LCR	55	60
AA-Omniscience	-10.5	-35.7
aider_polyglot	30.7	72
AIME 2025	11.7	92.7
AIR-Bench 2024	62.4	78.5
anthropic_red_team	99.1	98.2
arc_agi_1	4.5	58.7
ARC-AGI-2	0	6.1
arena_vision	1162	1201
Artificial Analysis Coding Index	24.2	25.6
bbq	95.1	94
critpt	0	0.6
Fortress	47.2	21.5
gdpval	0	25.4
GPQA Diamond	70.1	81.4
GSM8K	95.6	0
harmbench	82.9	97
HLE	5.3	16.5
IFBench	36	68.7
LiveCodeBench	38.3	84.5
livecodebench_easy	82.5	98.8
livecodebench_hard	4.5	62.9
livecodebench_medium	32.1	92.2
MathVista	63.8	84.3
mmlu_prox	61.1	69.3
MMMU	72.2	81.6
MMMU-Pro	59.9	69.2
multichallenge	40.3	44.9
MultiNRC	12.4	22.2
OmniScience Accuracy	23.7	24.8
OmniScience Non-Hallucination	62.1	19.5
PropensityBench	46.1	15.8
scicode	33.4	46.5
SEAL VISTA	34.9	51.8
simple_safety_tests	98.5	100
simplebench	17.8	38.7
swe_bench_bash	21.6	45
SWE-bench Verified	38.8	68.1
TAU-bench (airline)	42.8	49.2
TAU-bench (retail)	60.3	71.8
Terminal-Bench Hard	8.3	15.2
vectara_answer_rate	93.8	99.2
vectara_avg_summary_length	86.6	130.9
vectara_factual_consistency	90.4	81.4
vectara_hallucination_rate ↓	9.6	18.6
xstest	97.3	97.4
τ²-Bench Telecom (AA run)	28.9	55.6

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.