Which is better, Claude Sonnet 4 or GPT-5?

Across 60 shared benchmarks tracked on Vector Wire, Claude Sonnet 4 leads on 14 and GPT-5 leads on 46 — GPT-5 leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

Claude Sonnet 4 vs GPT-5

AnthropicvsOpenAI60 shared benchmarks14–46 head-to-head

Benchmark	Claude Sonnet 4	GPT-5
AA Agentic Index	39.2	49.7
AA Intelligence	29.8	35.3
AA-LCR	69.7	76.3
AA-Omniscience	0.2	-8.7
aider_polyglot	70.7	88
AIME	33.1	7.6
AIME 2025	74	95
AIR-Bench 2024	88.3	87.7
arc_agi_1	63.7	56.2
ARC-AGI-2	13.6	7.5
arena_vision	1207	1212
ArtifactsBench	57.3	73
Artificial Analysis Coding Index	37.6	38.9
browsecomp	12.2	54.9
browsecomp_zh	29.1	65
critpt	1.1	5.7
FinSearchComp-global	42	63.9
Fortress	24.4	17
frontiermath_tier_4	0	12.5
GAIA (text only)	68.3	76.4
gdpval	31.2	32.5
GPQA Diamond	78	87.3
harmbench	98	70
HLE	10.7	28.5
HLE (with tools)	20.3	35.2
HMMT 2025	15.9	93.3
IFBench	55	73.1
LiveCodeBench	68.5	85
LiveCodeBench v6	48.5	87
mmlu_redux	93.6	95.3
MMLU-Pro	84	87.5
MMMU	74.4	84.2
MMMU-Pro	62.4	78.4
Multi-SWE-Bench	35.7	39.3
multichallenge	57.1	71.1
MultiNRC	18.4	52.1
OmniScience Accuracy	22.7	40.3
OmniScience Non-Hallucination	70.9	21
scicode	40	43
SEAL VISTA	45.5	49.7
simplebench	45.5	56.7
swe_bench_bash	64.9	65
SWE-bench Multilingual	56.9	55.3
SWE-bench Pro	42.7	41.8
SWE-bench Verified	80.2	74.9
Tau2 airline	55.5	62.6
Tau2 retail	75	81.1
TauBench V3 - Banking	16.7	22.1
Terminal-Bench	36.4	43.8
Terminal-Bench 2.1	36.3	35.2
Terminal-Bench Hard	31.1	37.9
vectara_answer_rate	98.6	99.9
vectara_avg_summary_length	145.8	162.7
vectara_factual_consistency	89.7	85.3
vectara_hallucination_rate ↓	10.3	14.7
VisualToolBench	4.5	18.7
xbench_deepsearch	64.6	77.8
τ²-Bench	65	85
τ²-Bench Telecom (AA run)	64.6	86.5
τ³-Bench	13.8	19.6

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.