Which is better, GPT-5.4 or Qwen3.5 35B A3B?

Across 51 shared benchmarks tracked on Vector Wire, GPT-5.4 leads on 47 and Qwen3.5 35B A3B leads on 4 — GPT-5.4 leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

GPT-5.4 vs Qwen3.5 35B A3B

OpenAIvsAlibaba51 shared benchmarks47–4 head-to-head

Benchmark	GPT-5.4	Qwen3.5 35B A3B
AA Agentic Index	58.2	44.1
AA Intelligence	53.1	29.9
AA-LCR	77.7	68.3
AA-Omniscience	5.8	-48.1
AgentWorldBench - Android	60	53.2
AgentWorldBench - MCP	70.1	57.9
AgentWorldBench - OS	68.6	56.3
AgentWorldBench - Overall	58.3	47.7
AgentWorldBench - Search	37.3	26
AgentWorldBench - SWE	66.3	47.6
AgentWorldBench - Terminal	53.7	46.1
AgentWorldBench - Web	51.8	47.1
aime_2026	99.2	93.3
Artificial Analysis Coding Index	71.1	37
baby_vision	49.7	38.4
browsecomp	82.7	61
charxiv_rq	82.8	77.5
coding_arena_elo	1457	1250
critpt	23.4	0.9
gdpval	50.1	20.3
GPQA Diamond	93	84.5
HLE	43.7	47.4
hmmt_feb_2026	97.7	81.8
hmmt_nov_2025	95.8	89.2
IFBench	73.9	72.5
imo_answer_bench	91.4	76.8
itbenchSre	34.5	21.5
mathvision	92	83.9
MCP Atlas	70.6	62.4
mcpmark	62.5	27
MMLU-Pro	87.5	85.3
MMMU-Pro	81.2	75.1
nl2repo	41.3	20.5
OmniDocBench 1.5	89.1	89.3
OmniScience Accuracy	50.9	20.1
OmniScience Non-Hallucination	17.4	14.6
OSWorld-Verified	75	54.5
scicode	56.6	37.7
SWE-bench Multilingual	71.7	60.3
SWE-bench Pro	59.1	44.6
TauBench V3 - Banking	39.6	4.9
Terminal-Bench 2.0	75.1	40.5
Terminal-Bench 2.1	78.3	40.8
Terminal-Bench Hard	57.6	26.5
tool_decathlon	54.6	28.7
vectara_answer_rate	99.9	99.8
vectara_avg_summary_length	81.7	94.9
vectara_factual_consistency	93	89.5
vectara_hallucination_rate ↓	7	10.5
τ²-Bench Telecom (AA run)	87.1	89.2
τ³-Bench	72.9	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.