Inference
EN

MiniMax M3.1 Flash vs Step 5 Preview

Change models
MiniMax M3.1 FlashMiniMax
Step 5 PreviewStepFun
IndexAll categories, out of 100.
69.3 ±5.4 best
68.4 ±3.7
GapIndex points below the top.
Top
−0.9 · no clear gap
CoverageIndex weight that has results.
80%
90% best
SpeedOutput tokens per second.
100/s
105/s best
EloLMArena rating and rank.
—
1443
ReleasedMonth of release.
2026-10
2026-09
Input / 1MUS dollars per 1M input tokens.
$0.15 best
$1.00
Output / 1MUS dollars per 1M output tokens.
$0.60 best
$2.70
1K chats2K tokens in, 500 out, each.
$0.60 best
$3.35
1K documents20K tokens in, 1K out, each.
$3.60 best
$22.70
1K agent tasks200K tokens in, 20K out, each.
$42.00 best
$254.00
ContextMaximum tokens in one request.
1.05M best
1M
AgenticMulti-step tasks with tools.
71.8 best
68.8
CodingCode writing and repair.
69.4
71.4 best
ReasoningLogic problems and puzzles.
67.9 best
67.4
MultimodalTasks with images and text.
58.1
62.2 best
KnowledgeFacts and expert knowledge.
68.9 best
63.5
MultilingualTasks in many languages.
—
—
InstructionTasks with strict prompt rules.
—
—
MathMath problems.
—
61.7

Benchmarks

26 shared
Only MiniMax M3.1 Flash: 10Only Step 5 Preview: 36
MiniMax M3.1 Flash
Step 5 Preview
Artificial Analysis Humanity's Last ExamKnowledge
48.0% best17 / 222
46.5%23 / 222
Artificial Analysis Long Context ReasoningReasoning
79.7%45 / 217
88.3% best2 / 217
Critical Physics TasksReasoning
22.3% best25 / 214
20.9%26 / 214
Artificial Analysis Omniscience AccuracyKnowledge
45.1% best37 / 164
36.4%64 / 164
GDPval-AA normalizedAgentic
47.9%31 / 133
54.3% best24 / 133
Vibe Code Bench v1.1Coding
73.9% best45 / 123
70.4%48 / 123
Artificial Analysis SciCodeCoding
53.4%39 / 114
58.9% best13 / 114
Artificial Analysis MMMU-ProMultimodal
79.1% best29 / 113
76.4%44 / 113
Legal Research BenchLegal
34.1%50 / 90
40.4% best35 / 90
Tax Agent BenchFinance
32.1% best23 / 82
27.3%39 / 82
LiveBench Data AnalysisReasoning
71.4% best52 / 80
63.5%69 / 80
LiveBench ReasoningReasoning
82.9%48 / 80
83.8% best44 / 80
Terminal-Bench 4.0Agentic
25.3%28 / 61
31.8% best23 / 61
DeepSWEAgentic
73.7% best13 / 59
67.7%27 / 59
Agent Arena steerabilityAgentic
−2.636 / 54
−0.9 best30 / 54
Agent Arena task outcomeAgentic
−0.7 best34 / 54
−1.535 / 54
LiveBench Coding 2.0Coding
83.9%26 / 52
85.6% best22 / 52
LiveBench DesignAgentic
74.0%31 / 52
76.9% best24 / 52
LiveBench ResearchAgentic
72.7%29 / 52
72.8% best27 / 52
LiveBench Long AgenticAgentic
34.7% best29 / 51
27.5%36 / 51