Which is better, GPT-5.2 or GPT-5.4?

Across 69 shared benchmarks tracked on Vector Wire, GPT-5.2 leads on 14 and GPT-5.4 leads on 54 — GPT-5.4 leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

GPT-5.2 vs GPT-5.4

OpenAIvsOpenAI69 shared benchmarks14–54 head-to-head

Benchmark	GPT-5.2	GPT-5.4
AA Agentic Index	60.2	58.2
AA Intelligence	43.3	53.1
AA-LCR	79.3	77.7
AA-Omniscience	0.3	5.8
aime_2026	98.3	99.2
apexAgents	23	33.3
arc_agi_1	86.2	93.7
ARC-AGI-2	52.9	74
arena_elo	1437	1477
arena_text_factuality	1442	1476
arena_vision	1245	1277
Artificial Analysis Coding Index	48.7	71.1
browsecomp	65.8	82.7
browsecomp_with_context_manager	65.8	82.7
charxiv_rq	82.1	82.8
coding_arena_elo	1418	1457
critpt	11.6	23.4
deepsearchqa_f1	71.3	78.6
frontiermath_tier_4	18.8	27.1
gdpval	48.3	50.1
GDPval-AA (Elo)	1462	1674
GPQA Diamond	92.4	93
HealthBench Consensus length-adjusted	94.4	96.3
HealthBench Hard length-adjusted	34.3	29.1
HealthBench Professional length-adjusted	45.9	48.1
healthbench_length_adjusted	56.8	54
HLE	45.5	43.7
HLE (with tools)	45.5	52.1
hmmt_feb_2026	97	97.7
hmmt_nov_2025	99.2	95.8
IFBench	75.4	73.9
Image input eval - extremism (not_unsafe)	1	1
Image input eval - harms-erotic (not_unsafe)	1	1
Image input eval - hate (not_unsafe)	1	1
Image input eval - self-harm (not_unsafe)	1	1
imo_answer_bench	86.3	91.4
livebench	74.8	80.3
livebench_agentic_coding	50.3	53.8
livebench_coding	76.1	77.5
livebench_data_analysis	78.2	79.3
livebench_instruction_following	61.8	70.2
livebench_language	79.8	82.6
livebench_math	93.2	94.2
livebench_reasoning	83.2	88.1
matharena_visual_math_overall	86.5	92.5
mathvision	83	92
MCP Atlas	68	70.6
MMLU-Pro	87	87.5
MMMU-Pro	79.5	81.2
MultiNRC	42.2	58.3
OmniDocBench 1.5	85.7	89.1
OmniScience Accuracy	44.3	50.9
OmniScience Non-Hallucination	38.4	17.4
scicode	52.1	56.6
SEAL VISTA	46.6	50.9
simpleqa_verified	38.9	45.3
SWE-bench Multilingual	72	71.7
SWE-bench Pro	55.6	59.1
TauBench V3 - Banking	11.1	39.6
Terminal-Bench 2.0	54	75.1
Terminal-Bench Hard	62.2	57.6
tool_decathlon	46.3	54.6
toolathlon	46.3	54.6
vectara_answer_rate	100	99.9
vectara_avg_summary_length	186.3	81.7
vectara_factual_consistency	89.2	93
vectara_hallucination_rate ↓	10.8	7
τ²-Bench Telecom (AA run)	98.7	87.1
τ³-Bench	12	72.9

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.