Which is better, Claude Opus 4.5 or Qwen3.5 27B?

Across 69 shared benchmarks tracked on Vector Wire, Claude Opus 4.5 leads on 47 and Qwen3.5 27B leads on 22 — 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 Qwen3.5 27B

AnthropicvsAlibaba69 shared benchmarks47–22 head-to-head

Benchmark	Claude Opus 4.5	Qwen3.5 27B
AA Agentic Index	59.6	54.6
AA Intelligence	41.9	34.6
AA-LCR	76	72.3
AA-Omniscience	14	-44
aime_2026	95.1	92.6
Artificial Analysis Coding Index	47.8	34.9
browsecomp	37	61
browsecomp_zh	62.4	62.1
C-Eval	92.2	90.5
CC-OCR	76.9	81
charxiv_rq	68.5	79.5
Claw Eval (pass@3)	59.6	46.2
Claw-Eval Avg	76.6	64.3
coding_arena_elo	1495	1357
critpt	4.6	0.9
DynaMath	79.7	87.7
ERQA	46.8	60.5
gdpval	47.3	33
GPQA Diamond	87	85.8
HLE	30.8	48.5
hmmt_feb_2025	92.9	92
hmmt_feb_2026	85.3	84.3
hmmt_nov_2025	93.3	89.8
IFBench	58	76.5
imo_answer_bench	84	79.9
LiveCodeBench v6	84.8	80.7
longbench_v2	64.4	60.6
mathvision	77.1	86
MathVista	80.2	87.8
MCP Atlas	69.8	68.4
MLVU	81.7	85.9
mmlu_redux	95.6	93.2
MMLU-Pro	90	86.1
mmmlu	90.8	85.9
MMMU	80.7	82.3
MMMU-Pro	74	75
MMStar	73.2	81
MMVU	77.3	73.3
multichallenge	59	60.8
MV-Bench	67.2	74.6
nl2repo	43.2	27.3
OCRBench	86.5	89.4
OmniDocBench 1.5	87.7	88.9
OmniScience Accuracy	46.6	20.7
OmniScience Non-Hallucination	39	24.9
OSWorld-Verified	66.3	56.2
QwenClawBench	52.3	52.2
QwenWebBench	1536	1068
RealWorldQA	77	83.7
scicode	50	39.5
Seal-0	47.7	47.2
SimpleVQA	69.7	56
SkillsBench Avg5	45.3	27.2
supergpqa	70.6	65.6
SWE-bench Multilingual	77.5	69.3
SWE-bench Pro	57.1	51.2
SWE-bench Verified	81.5	75
Terminal-Bench 2.0	59.3	41.6
Terminal-Bench Hard	47	32.6
tool_decathlon	43.5	31.5
V*	67	93.7
vectara_answer_rate	98.7	99.8
vectara_avg_summary_length	114.5	94.4
vectara_factual_consistency	89.1	87.9
vectara_hallucination_rate ↓	10.9	12.1
VideoMME (w sub.)	77.7	87
VideoMMMU	84.4	82.3
ZeroBench	3	0.1
τ²-Bench Telecom (AA run)	89.5	93.9

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.