priorsBuilder

MiniMax · released 2026-06-01

MiniMax M3

Measured on 80 benchmarks across 4 reasoning efforts.

General frontier

65.0

#12 of 136 models · best at none

includes estimates: 5% of the mix is measured for this model

Reasoning efforts in your mix

none65.0#28 of 313
xhigh63.4#36 of 313
default41.5#131 of 313
max25.2#248 of 313

Your mix’s benchmarks

4 of 8 measured

reasoning

Humanity's Last Exam

587 results
  1. #93default of 58739.0%

Leader: Opus 5.5 · 61.4%

knowledge

AA-Omniscience

526 results
  1. #120default of 5261.4

Leader: Opus 5.5 · 46.4

reasoning

BullshitBench v2

196 results
  1. #40xhigh of 19664.0%
  2. #44none of 19663.0%

Leader: Opus 4.8 · 95.0%

reasoning

SimpleBench

97 results
  1. #60max of 9745.8%

Leader: Opus 5.5 · 88.4%

All 80 benchmarks

show 76 more

agentic

AA-AnalystAgent

44 results
  1. #39default of 4410.0%

Leader: Gemini 3.7 Flash · 60.0%

agentic

AA-Briefcase v1.1

224 results
  1. #103default of 2241091

Leader: Sonnet 5.5 · 1823

agentic

Agent Arena

50 results
  1. #43max of 50-7.0%

Leader: Opus 5.5 · 15.5%

agentic

AutomationBench (Zapier)

115 results
  1. #93max of 1154.9%

Leader: Gemini 4 · 51.3%

agentic

AutomationBench-AA

219 results
  1. #129default of 21921.3%

Leader: Gemini 4 · 77.5%

agentic

EnterpriseOps-Gym-AA

50 results
  1. #39default of 5032.1%

Leader: Fable 5 · 51.1%

agentic

Harvey's Legal Agent Benchmark

79 results
  1. #38max of 794.2%

Leader: Muse Spark 1.2 · 25.4%

agentic

Legal Research Bench

78 results
  1. #47max of 7829.8%

Leader: Muse Spark 1.3 · 55.3%

agentic

OSWorld-Verified

13 results
  1. #4max of 1375.2%

Leader: Fable 5 · 86.0%

agentic

SkillsBench

34 results
  1. #19max of 3451.5%

Leader: DeepSeek V4.1 Flash · 69.8%

agentic

Tax Agent Bench

69 results
  1. #45max of 6920.6%

Leader: Fable 5.1 · 49.2%

agentic

Terminal-Bench 4.0 (Vals)

47 results
  1. #43max of 471.0%

Leader: Opus 5.5 · 65.2%

agentic

Vending-Bench 2

67 results
  1. #50max of 67$2,158

Leader: GPT-6 Astra · $15,515

agentic

τ²-Bench Telecom

393 results
  1. #71default of 39388.9%

Leader: GLM 5.2 · 99.1%

agentic

τ³-Banking

206 results
  1. #118default of 20615.3%

Leader: Qwen 3.8 Max · 51.3%

coding

Arena Image-to-WebDev

58 results
  1. #43max of 581473

Leader: Opus 5.5 · 1749

coding

Arena WebDev

122 results
  1. #60max of 1221482

Leader: Opus 5.5 · 1813

coding

Code Migration

77 results
  1. #52max of 7719.9%

Leader: Sonnet 5.5 · 69.8%

coding

DeepSWE v1

29 results
  1. #19max of 2920.4%

Leader: GPT-5.5 · 70.0%

coding

FrontierCode 1.1 Extended

130 results
  1. #124none of 13028.5%

Leader: Opus 5.5 · 65.3%

coding

FrontierCode 1.1 Main

130 results
  1. #124none of 13014.7%

Leader: Opus 5.5 · 54.6%

coding

LiveCodeBench (Vals)

127 results
  1. #56max of 12782.2%

Leader: Fable 5.1 · 90.5%

coding

SciCode

226 results
  1. #128default of 22647.1%

Leader: Opus 5.5 · 66.9%

coding

SWE-bench Verified (Vals)

85 results
  1. #43max of 8575.0%

Leader: Opus 5 · 97.0%

coding

Terminal-Bench 2.0 (Vals)

63 results
  1. #22max of 6346.1%

Leader: GPT-5.5 · 73.2%

coding

Terminal-Bench 2.1 (Vals)

74 results
  1. #53max of 7453.6%

Leader: Opus 5.5 · 87.6%

coding

Terminal-Bench Hard-AA

387 results
  1. #46default of 38742.4%

Leader: GPT-5.6 Sol · 65.9%

coding

Terminal-Bench Science-AA

46 results
  1. #41default of 460.5%

Leader: GPT-6 Astra · 63.3%

coding

Terminal-Bench v2.1-AA

238 results
  1. #81default of 23865.2%

Leader: Fable 5.1 · 91.4%

coding

Terminal-Bench v4-AA

216 results
  1. #113default of 2162.0%

Leader: Sonnet 5.5 · 63.6%

coding

Vibe Code Bench 1-100

22 results
  1. #20max of 229.2%

Leader: Opus 5.5 · 30.4%

coding

Vibe Code Bench v1.1

105 results
  1. #55max of 10547.6%

Leader: Sonnet 5.5 · 92.4%

composite

AA Intelligence Index v4.3.2

610 results
  1. #125default of 61029.2

Leader: Opus 5.5 · 57.6

composite

Arena Text

220 results
  1. #81max of 2201440

Leader: Gemini 4 · 1525

composite

Arena Text — Multi-Turn

220 results
  1. #75max of 2201452

Leader: Gemini 4 · 1553

composite

Vals Index

47 results
  1. #41max of 4736.5%

Leader: Gemini 4 · 68.9%

composite

Vals Multimodal Index v1.2

33 results
  1. #18max of 3360.0%

Leader: Fable 5 · 74.2%

cost

AA Intelligence Index Cost per Task

177 results
  1. #90default of 177$0.51

Leader: GPT-6 Luna · $0.00

cost

Cost to Run AA Intelligence Index

177 results
  1. #73default of 177$538

Leader: GPT-6 Luna · $10.63

creative

VoxelBench (text)

55 results
  1. #42max of 551344

Leader: GPT-6 Astra · 2732

economics

CorpFin (Vals)

117 results
  1. #11max of 11768.1%

Leader: Opus 5 · 73.2%

economics

Excel Modeling Benchmark

74 results
  1. #55max of 7447.8%

Leader: Fable 5.1 · 76.7%

economics

Finance Agent v2

78 results
  1. #49max of 7848.3%

Leader: Gemini 4 · 65.4%

economics

GDP.pdf

210 results
  1. #131default of 2109.8%

Leader: Muse Wonder Full (Max) · 32.6%

economics

GDPval-AA v2.1

288 results
  1. #108default of 2881242

Leader: Opus 5.5 · 1866

economics

MortgageTax (Vals)

84 results
  1. #16max of 8468.4%

Leader: Opus 5 · 72.1%

economics

Public Benefits Bench v1 (Vals)

13 results
  1. #3max of 1360.7%

Leader: Fable 5 · 71.7%

economics

TaxEval (Vals)

125 results
  1. #58max of 12572.7%

Leader: Muse Spark 1.2 · 80.4%

instruction

IFBench

399 results
  1. #3default of 39982.9%

Leader: Grok 4.3 · 83.3%

knowledge

AA-Omniscience Accuracy

526 results
  1. #362default of 52616.7%

Leader: Fable 5.1 · 67.2%

knowledge

AA-Omniscience Non-hallucination

526 results
  1. #12default of 52681.6%

Leader: MiniCPM5-1B (Non-reasoning) · 99.1%

knowledge

GPQA Diamond

539 results
  1. #27default of 53992.9%

Leader: GPT-6 Astra · 96.3%

knowledge

GPQA Diamond (Vals)

122 results
  1. #15max of 12292.7%

Leader: Gemini 3.1 Pro · 95.5%

knowledge

Harvey LAB-AA v1.1

27 results
  1. #24default of 270.0%

Leader: Muse Wonder Full (Max) · 13.1%

knowledge

MedCode

94 results
  1. #35max of 9446.3%

Leader: Opus 5 · 63.6%

knowledge

MedScribe

96 results
  1. #17max of 9687.3%

Leader: Opus 5.5 · 91.4%

knowledge

MMLU Pro (Vals)

122 results
  1. #63max of 12284.2%

Leader: Fable 5.1 · 92.4%

knowledge

MMMU Pro (Vals)

81 results
  1. #41max of 8181.2%

Leader: Fable 5.1 · 90.6%

knowledge

MMMU-Pro

265 results
  1. #67default of 26578.6%

Leader: Opus 5.5 · 87.7%

knowledge

Public Benefits Bench v1.1

50 results
  1. #24max of 5064.1%

Leader: Opus 5 · 76.9%

long-context

AA-LCR v1.1

523 results
  1. #33default of 52383.0%

Leader: Kimi K3 · 88.7%

long-context

Arena Document

43 results
  1. #31max of 431435

Leader: Opus 5 · 1516

long-context

MLCR-AA

94 results
  1. #40default of 9417.2%

Leader: Sonnet 5.5 · 75.0%

math

ProofBench v1.1

52 results
  1. #43max of 5218.0%

Leader: Sonnet 5.5 · 100.0%

other

Arena Vision

116 results
  1. #62max of 1161236

Leader: Fable 5 · 1308

reasoning

CritPt

529 results
  1. #166default of 5293.7%

Leader: Muse Wonder Full (Max) · 32.9%

reasoning

LegalBench

135 results
  1. #20max of 13585.4%

Leader: Fable 5 · 88.6%

reasoning

SAGE

80 results
  1. #16max of 8050.6%

Leader: Opus 4.7 · 56.1%

reliability

Tool-call error rate

171 results
  1. #128max of 1713.5%

Leader: GPT-5.4 Pro · 0.0%

reliability

Uptime (7d)

181 results
  1. #153max of 18197.8%

Leader: GPT-5 Pro · 100.0%

safety

CWE-Bench

19 results
  1. #16default of 1941.7%

Leader: Grok 4.7 · 68.3%

safety

CyberBench v1.1

45 results
  1. #33max of 4565.0%

Leader: Wonder · 80.4%

speed

Latency (Time to First Token)

351 results
  1. #58default of 3510.9s

Leader: Gemini 2.5 Flash-Lite · 0.3s

speed

Output Speed (tokens/s)

351 results
  1. #146default of 351108.3

Leader: Celeris-1 · 1522.5

speed

Provider latency

177 results
  1. #87max of 1771.3s

Leader: Qwen 2.5 7B · 0.2s

speed

Provider throughput

176 results
  1. #70max of 17673.8

Leader: Muse Spark 1.1 · 204.3