Which is better, Claude Sonnet 4.6 or Qwen3.5 35B A3B?

Across 42 shared benchmarks tracked on Vector Wire, Claude Sonnet 4.6 leads on 37 and Qwen3.5 35B A3B leads on 5 — Claude Sonnet 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 Sonnet 4.6 vs Qwen3.5 35B A3B

AnthropicvsAlibaba42 shared benchmarks37–5 head-to-head

Benchmark	Claude Sonnet 4.6	Qwen3.5 35B A3B
AA Agentic Index	61.6	44.1
AA Intelligence	48.4	29.9
AA-LCR	74	68.3
AA-Omniscience	12.2	-48.1
AgentWorldBench - Android	58	53.2
AgentWorldBench - MCP	70	57.9
AgentWorldBench - OS	63.2	56.3
AgentWorldBench - Overall	56	47.7
AgentWorldBench - Search	28.8	26
AgentWorldBench - SWE	64.5	47.6
AgentWorldBench - Terminal	57	46.1
AgentWorldBench - Web	50.8	47.1
Artificial Analysis Coding Index	63	37
browsecomp	76.2	61
charxiv_rq	72.4	77.5
coding_arena_elo	1523	1250
critpt	3.1	0.9
gdpval	54.8	20.3
GPQA (unspecified)	89.9	84.2
GPQA Diamond	89.9	84.5
HLE	49	47.4
IFBench	56.6	72.5
itbenchSre	39.8	21.5
MCP Atlas	69.5	62.4
mmmlu	89.3	85.2
MMMU-Pro	75.6	75.1
OmniScience Accuracy	40.9	20.1
OmniScience Non-Hallucination	51.6	14.6
OSWorld-Verified	78.5	54.5
scicode	47	37.7
SWE-bench Pro	58.1	44.6
SWE-bench Verified	79.6	70
TauBench V3 - Banking	34.4	4.9
Terminal-Bench 2.0	59.1	40.5
Terminal-Bench 2.1	71.2	40.8
Terminal-Bench Hard	59.1	26.5
vectara_answer_rate	99.9	99.8
vectara_avg_summary_length	114.7	94.9
vectara_factual_consistency	89.4	89.5
vectara_hallucination_rate ↓	10.6	10.5
τ²-Bench Telecom (AA run)	97.9	89.2
τ³-Bench	30.5	68.9

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.