Atlas

Compare Models

Gemini 4 Carbon vs Kimi K3.1 Pro

Gemini 4 Carbon leads by 4.1 capability points (≈95% interval +1.5 to +6.7): a clear lead over Kimi K3.1 Pro · P(higher) ≈ 100%

Wins 46 of 53 shared benchmarks (7 lost)

Capability vs. relative cost
Gemini 4 CarbonKimi K3.1 ProClosed weightsOpen weights

Each line runs through one model’s priced reasoning settings, cheapest to most expensive; grey points are other models’ priced settings in the same range. Results whose reasoning effort the source didn’t report are left out wherever a model has named effort levels. Capability and relative cost are the published estimates (relative cost is cost per task as a multiple of GPT-5 (high)); hover a setting, or use the arrow keys, to set it beside the other model’s same setting.

Gemini 4 CarbonvsKimi K3.1 Procapability vs. relative costData Nov 15, 2026Village LLM Atlasatlas.village.dev

Benchmark wins46 – 7 · 53 shared

Comparing matching setups only (38 benchmarks excluded) · suites counted once
← Gemini 4 CarbonKimi K3.1 Pro →

Gemini 4 CarbonvsKimi K3.1 Prowins by categoryData Nov 15, 2026Data: Vals AI, Arena, MathArena and 11 othersVillage LLM Atlasatlas.village.dev

Margins

Each benchmark scaled 0–100% (its declared range, else the field’s range); Δ is Kimi K3.1 Pro minus Gemini 4 Carbon.

ArXivMath 10/2026 (Native Harnesses)Math93.645.7−47.9
ArXivLean 09/2026Math62.317.0−45.3
ExploitGym - release-reported percentage, cohort unspecifiedCybersecurity54.820.7−34.1
ArXivMath 09/2026 (Native Harnesses)Math96.264.2−32.1
Legal Research Bench FamilyLegal45.513.6−31.8
Legal Research Bench CriminalLegal59.628.1−31.6
Humanity's Last Exam DiamondGeneral QA69.139.4−29.7
ArXivMath 08/2026 (Native Harnesses)Math93.064.9−28.1
Legal Research Bench ImmigrationLegal45.518.2−27.3
FrontierCode 2.0 - Pass RateCode46.922.0−25.0
BrokenArXiv 10/2026 (Native Harnesses)Math82.758.0−24.7
BrokenArXiv 08/2026 (Native Harnesses)Math88.164.3−23.8
Legal Research Bench Constitutional/Civil RightsLegal55.331.6−23.7
LMArena Agent Arena Net Improvement (Steering Burden)Agents17.04.9−22.8
FrontierCode 2.0 - ScoreCode42.619.9−22.7
Code Migration v1.1 (Vals Index v2.2 subset)Code80.159.7−20.4
Bug Hunt BenchCode54.634.3−20.3
Legal Research Bench Weighted PassLegal91.671.4−20.2
RiemannBench (November 2026 rescoring)Math80.060.8−19.2
OpenWorldEngineer (Mean@8)Code32.813.6−19.2
BrokenArXiv 09/2026 (Native Harnesses)Math82.463.5−18.9
Legal Research Bench Business/CommercialLegal46.527.9−18.6
Legal Research Bench Civil LitigationLegal48.129.6−18.5
LMArena Agent Arena Praise vs ComplaintAgents41.08.0−16.5
APEX-Agents 1.1Professional Work90.475.9−14.5
FrontierBioMLScience39.025.8−13.1
Tax Agent BenchFinance & Accounting76.463.9−12.5
Vals Finance IndexIndexes67.056.0−11.0
APEX-1 (Mercor, mean score)Professional Work80.069.3−10.7
LMArena Code Arena WebDevCode17681688−10.5
BuseyBench SVGMultimodal7.96.9−9.2
Harvey LAB (Mercor run, undisclosed ~213-task cohort)Legal23.013.8−9.2
Legal Research Bench Administrative/RegulatoryLegal65.256.1−9.1
APEX-Accounting 1.1Finance & Accounting70.762.4−8.3
LMArena Agent Arena Confirmed SuccessAgents18.42.7−7.8
LMArena Text Arena EloChat & Writing15261483−7.5
DAXBench v2Code93.686.2−7.4
Artificial Analysis Coding Agent Index v2.1Indexes55.548.6−6.8
Excel Modeling Benchmark v1.1Finance & Accounting76.570.0−6.5
LMArena Agent Arena Bash ReliabilityAgents-5.27.2+6.2
LMArena Agent Arena Steering BurdenAgents19.16.7−6.2
HLAB-Vals-1.1Legal17.511.7−5.8
Finance Agent v2.1Finance & Accounting67.261.5−5.7
MMMU ProMultimodal89.284.1−5.1
GDPval-AA v2.1Professional Work17331846+5.0
AutomationBench-AA 1.0.6Agents77.873.3−4.5
Vibe Code Bench v1.1Code89.184.9−4.2
τ³-Banking (Mercor run, bm25_grep retrieval)Agents55.759.8+4.1
AA-Briefcase v1.1 EloAgents16111678+3.7
Legal Research Bench HealthLegal70.073.3+3.3
Agents' Last Exam v1.1 ScoreAgents77.474.4−3.0
Terminal-Bench 4.0Code63.163.6+0.5
LMArena Agent Arena Tool HallucinationAgents0.51.0+0.2

Gemini 4 CarbonvsKimi K3.1 Pronormalised score marginsData Nov 15, 2026Data: Vals AI, Arena, MathArena and 11 othersVillage LLM Atlasatlas.village.dev

Win counts use each benchmark’s declared score direction. Represented suite slices and the derived Capability score are excluded from headline totals. Missing or incompatible setups are excluded, never counted as losses.
Loading Atlas data…