Which is better, GPT-5.1 or GPT-5.4?

Across 48 shared benchmarks tracked on Vector Wire, GPT-5.1 leads on 5 and GPT-5.4 leads on 43 — 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-5.1 vs GPT-5.4

OpenAIvsOpenAI48 shared benchmarks5–43 head-to-head

Benchmark	GPT-5.1	GPT-5.4
AA Agentic Index	32.2	58.2
AA Intelligence	37.5	53.1
AA-LCR	76.7	77.7
AA-Omniscience	5.4	5.8
arc_agi_1	72.8	93.7
ARC-AGI-2	17.6	74
arena_elo	1439	1477
arena_text_factuality	1445	1476
arena_vision	1250	1277
Artificial Analysis Coding Index	49.4	71.1
browsecomp	50.8	82.7
coding_arena_elo	1421	1457
critpt	4.9	23.4
frontiermath_tier_4	12.5	27.1
gdpval	25	50.1
GPQA Diamond	88.1	93
HealthBench Consensus length-adjusted	95	96.3
HealthBench Hard length-adjusted	25.4	29.1
HealthBench Professional length-adjusted	39.6	48.1
healthbench_length_adjusted	50.9	54
HLE	28.5	43.7
HLE (with tools)	42.7	52.1
hmmt_nov_2025	91.7	95.8
IFBench	72.9	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
matharena_visual_math_overall	76.9	92.5
MCP Atlas	50.1	70.6
MMLU-Pro	87	87.5
MMMU-Pro	75.5	81.2
MultiNRC	49	58.3
OmniScience Accuracy	37.7	50.9
OmniScience Non-Hallucination	48.1	17.4
scicode	43.3	56.6
SEAL VISTA	43.8	50.9
simpleqa_verified	48.9	45.3
TauBench V3 - Banking	15.9	39.6
Terminal-Bench 2.0	47.6	75.1
Terminal-Bench 2.1	52.4	78.3
Terminal-Bench Hard	45.5	57.6
vectara_answer_rate	100	99.9
vectara_avg_summary_length	254.4	81.7
vectara_factual_consistency	89.1	93
vectara_hallucination_rate ↓	12.1	7
τ²-Bench Telecom (AA run)	81.9	87.1
τ³-Bench	14	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.