Which is better, Claude Opus 4.6 or Claude Sonnet 4.5?

Across 60 shared benchmarks tracked on Vector Wire, Claude Opus 4.6 leads on 52 and Claude Sonnet 4.5 leads on 8 — Claude Opus 4.6 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 4.6 vs Claude Sonnet 4.5

AnthropicvsAnthropic60 shared benchmarks52–8 head-to-head

Benchmark	Claude Opus 4.6	Claude Sonnet 4.5
AA Agentic Index	67.6	50.6
AA Intelligence	44.9	37.4
AA-LCR	74.3	68.3
AA-Omniscience	13.7	-0.1
AIME 2025	99.8	88
AIME25 no tools	95.6	88
arc_agi_1	93	63.7
ARC-AGI-2	68.8	13.6
arena_elo	1497	1456
arena_text_factuality	1487	1448
Artificial Analysis Coding Index	48.1	52.1
browsecomp	86.8	24.1
charxiv_rq	69.1	67.2
coding_arena_elo	1545	1392
critpt	12.6	1.1
Fortress	20.5	17
frontiermath_tier_4	22.9	4.2
gdpval	55.9	40.4
GPQA Diamond	91.3	83.4
HLE	53.1	19.8
HLE (with tools)	62.7	32
hmmt_nov_2025	96.3	81.7
IFBench	62.5	57.3
imo_answer_bench	75.3	65.9
LiveCodeBench	88.8	71
LiveCodeBench v6	88.8	64
matharena_visual_math_overall	72.3	75.8
MCP Atlas	76.8	59.5
MHGRN (8-needle 1M variant)	76	18.5
MMLU-Pro	89.1	88.2
mmmlu	91.1	89.1
MMMU-Pro	77.3	68.7
multichallenge	56	55.3
MultiNRC	57.1	35.8
Needle-in-a-Haystack (8-needle 1M variant)	76	18.5
OmniDocBench 1.5	86.6	85.8
OmniScience Accuracy	47	32.9
OmniScience Non-Hallucination	37.2	50.9
OSWorld-Verified	72.7	61.4
RealWorldQA	73.9	70.3
scicode	52	45
SEAL VISTA	46.1	48.8
simplebench	67.6	54.3
simpleqa_verified	46.5	23.6
swe_bench_bash	75.6	71.4
SWE-bench Multilingual	77.8	68
SWE-bench Pro	57.3	43.6
SWE-bench Verified	80.8	82
Terminal-Bench 2.0	65.4	50
Terminal-Bench Hard	65.4	35.6
toolathlon	56.8	41
Toolathlon Avg turns	16.9	32
Toolathlon Pass@∞	47.2	28.7
vectara_answer_rate	99.8	95.6
vectara_avg_summary_length	137.6	127.8
vectara_factual_consistency	87.8	88
vectara_hallucination_rate ↓	12.2	12
VisualToolBench	27.5	6.2
τ²-Bench Telecom (AA run)	99.3	78.1
τ³-Bench	72.4	19

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.