Which is better, GPT-5.4 or Qwen3.6 Plus?

Across 57 shared benchmarks tracked on Vector Wire, GPT-5.4 leads on 47 and Qwen3.6 Plus leads on 10 — 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 Qwen3.6 Plus

OpenAIvsAlibaba57 shared benchmarks47–10 head-to-head

Benchmark	GPT-5.4	Qwen3.6 Plus
AA Agentic Index	58.2	29
AA Intelligence	53.1	40.5
AA-LCR	77.7	72.3
AA-Omniscience	5.8	2.6
AgentWorldBench - Android	60	57.6
AgentWorldBench - MCP	70.1	55.3
AgentWorldBench - OS	68.6	60.3
AgentWorldBench - Overall	58.3	50.8
AgentWorldBench - Search	37.3	21.9
AgentWorldBench - SWE	66.3	59.1
AgentWorldBench - Terminal	53.7	50.6
AgentWorldBench - Web	51.8	50.8
aime_2026	99.2	95.3
arena_elo	1477	1444
arena_text_factuality	1476	1438
Artificial Analysis Coding Index	71.1	54.5
coding_arena_elo	1457	1460
critpt	23.4	2.9
frontiermath_tier_4	27.1	8.3
gdpval	50.1	32
GPQA Diamond	93	90.4
HLE	43.7	28.8
HLE (with tools)	52.1	50.6
hmmt_feb_2026	97.7	87.8
hmmt_nov_2025	95.8	94.6
IFBench	73.9	75.2
imo_answer_bench	91.4	83.8
livebench	80.3	70.8
livebench_agentic_coding	53.8	41.4
livebench_coding	77.5	78.2
livebench_data_analysis	79.3	69.9
livebench_instruction_following	70.2	58.3
livebench_language	82.6	75
livebench_math	94.2	83.7
livebench_reasoning	88.1	75.8
mathvision	92	88
MCP Atlas	70.6	74.1
mcpmark	62.5	48.2
MMLU-Pro	87.5	88.5
MMMU-Pro	81.2	78.8
nl2repo	41.3	37.9
OmniDocBench 1.5	89.1	91.2
OmniScience Accuracy	50.9	26.4
OmniScience Non-Hallucination	17.4	68
OSWorld-Verified	75	62.5
scicode	56.6	40.7
simpleqa_verified	45.3	49.1
SWE-bench Multilingual	71.7	73.8
SWE-bench Pro	59.1	56.6
TauBench V3 - Banking	39.6	20.8
Terminal-Bench 2.0	75.1	61.6
Terminal-Bench 2.1	78.3	61.4
Terminal-Bench Hard	57.6	43.9
tool_decathlon	54.6	39.8
toolathlon	54.6	39.8
τ²-Bench Telecom (AA run)	87.1	97.7
τ³-Bench	72.9	70.7

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.