Inference
EN

GPT-5.4 vs Qwen3.8-Flash-Next

Change models
GPT-5.4OpenAI
Qwen3.8-Flash-NextAlibaba
IndexAll categories, out of 100.
64.8 ±2.5 best
64.3 ±3.2
GapIndex points below the top.
Top
−0.5 · no clear gap
CoverageIndex weight that has results.
95%
95%
SpeedOutput tokens per second.
38/s
55/s best
EloLMArena rating and rank.
1452
—
ReleasedMonth of release.
2026-03
2026-08
Input / 1MUS dollars per 1M input tokens.
$2.50
—
Output / 1MUS dollars per 1M output tokens.
$15.00
—
Batch output / 1MBatch price per 1M output tokens.
$7.50
—
1K chats2K tokens in, 500 out, each.
$12.50
—
1K documents20K tokens in, 1K out, each.
$65.00
—
1K agent tasks200K tokens in, 20K out, each.
$800.00
—
ContextMaximum tokens in one request.
1.05M best
262K
AgenticMulti-step tasks with tools.
62.1
68.5 best
CodingCode writing and repair.
61.8
62.1 best
ReasoningLogic problems and puzzles.
69.0 best
62.2
MultimodalTasks with images and text.
62.1
65.0 best
KnowledgeFacts and expert knowledge.
66.4 best
55.2
MultilingualTasks in many languages.
—
—
InstructionTasks with strict prompt rules.
63.4
73.8 best
MathMath problems.
68.0 best
59.4

Benchmarks

24 shared
Only GPT-5.4: 75Only Qwen3.8-Flash-Next: 25
GPT-5.4
Qwen3.8-Flash-Next
Artificial Analysis Humanity's Last ExamKnowledge
43.7% best31 / 222
38.0%55 / 222
Artificial Analysis Long Context ReasoningReasoning
82.0% best29 / 217
79.7%45 / 217
Critical Physics TasksReasoning
23.4% best23 / 214
11.1%61 / 214
Artificial Analysis GPQA DiamondKnowledge
92.0%22 / 199
92.3% best19 / 199
Artificial Analysis Omniscience AccuracyKnowledge
50.8% best18 / 164
23.5%112 / 164
Graduate-Level Google-Proof Q&AKnowledge
92.8% best13 / 141
91.7%21 / 141
GDPval-AA normalizedAgentic
37.3%47 / 133
56.4% best13 / 133
Artificial Analysis MMMU-ProMultimodal
78.4%37 / 113
79.8% best24 / 113
GPQA Diamond · BenchLMKnowledge
92.8% best14 / 84
91.7%21 / 84
SWE-bench ProCoding
57.7%41 / 83
62.5% best24 / 83
LiveBench Data AnalysisReasoning
79.3% best15 / 80
74.2%43 / 80
LiveBench ReasoningReasoning
88.1% best25 / 80
87.4%28 / 80
Humanity's Last ExamKnowledge
52.1% best20 / 77
35.9%39 / 77
LiveBench Agentic CodingAgentic
53.8%35 / 66
61.6% best12 / 66
LiveBench CodingCoding
77.5% best35 / 66
72.6%53 / 66
LiveBench Instruction FollowingInstruction
70.2%27 / 66
77.1% best5 / 66
LiveBench LanguageKnowledge
82.6% best23 / 66
74.6%49 / 66
LiveBench MathematicsMath
94.1% best17 / 66
85.8%50 / 66
Humanity's Last Exam without toolsKnowledge
39.8% best29 / 53
35.9%32 / 53
Agent Arena command recoveryAgentic
1.2 best26 / 50
−2.639 / 50