Which is better, Claude Opus 4.5 or Claude Sonnet 4?

Across 48 shared benchmarks tracked on Vector Wire, Claude Opus 4.5 leads on 43 and Claude Sonnet 4 leads on 5 — Claude Opus 4.5 leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

Claude Opus 4.5 vs Claude Sonnet 4

AnthropicvsAnthropic48 shared benchmarks43–5 head-to-head

Benchmark	Claude Opus 4.5	Claude Sonnet 4
AA Agentic Index	59.6	39.2
AA Intelligence	41.9	29.8
AA-LCR	76	69.7
AA-Omniscience	14	0.2
AIME 2025	92.8	74
arc_agi_1	82	63.7
ARC-AGI-2	37.6	13.6
Artificial Analysis Coding Index	47.8	37.6
browsecomp	37	12.2
browsecomp_zh	62.4	29.1
critpt	4.6	1.1
Fortress	13.6	24.4
frontiermath_tier_4	4.2	0
gdpval	47.3	31.2
GPQA Diamond	87	78
HLE	30.8	10.7
HLE (with tools)	43.4	20.3
IFBench	58	55
livebench	76	74.8
LiveCodeBench	87	68.5
LiveCodeBench v6	84.8	48.5
mmlu_redux	95.6	93.6
MMLU-Pro	90	84
mmmlu	90.8	86.5
MMMU	80.7	74.4
MMMU-Pro	74	62.4
Multi-SWE-Bench	50	35.7
multichallenge	59	57.1
MultiNRC	48.6	18.4
OmniScience Accuracy	46.6	22.7
OmniScience Non-Hallucination	39	70.9
OSWorld-Verified	66.3	42.2
scicode	50	40
SEAL VISTA	46.4	45.5
simplebench	62	45.5
supergpqa	70.6	55.7
swe_bench_bash	76.8	64.9
SWE-bench Multilingual	77.5	56.9
SWE-bench Pro	57.1	42.7
SWE-bench Verified	81.5	80.2
Tau2 retail	88.9	75
Terminal-Bench Hard	47	31.1
vectara_answer_rate	98.7	98.6
vectara_avg_summary_length	114.5	145.8
vectara_factual_consistency	89.1	89.7
vectara_hallucination_rate ↓	10.9	10.3
τ²-Bench	98.2	65
τ²-Bench Telecom (AA run)	89.5	64.6

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.