Atlas

Compare Models

Claude Opus 5 vs Gemini 4 Carbon

Gemini 4 Carbon leads by 3.9 capability points (≈95% interval +1.2 to +6.5): a clear lead over Claude Opus 5 · P(higher) ≈ 100%

Wins 45 of 50 shared benchmarks (5 lost)

Capability vs. relative cost
Claude Opus 5Gemini 4 CarbonClosed 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.

Claude Opus 5vsGemini 4 Carboncapability vs. relative costData Nov 15, 2026Village LLM Atlasatlas.village.dev

Benchmark wins5 – 45 · 50 shared

Comparing matching setups only (3 benchmarks excluded) · suites counted once
← Claude Opus 5Gemini 4 Carbon →

Claude Opus 5vsGemini 4 Carbonwins by categoryData Nov 15, 2026Data: Artificial Analysis, Arena, Vals AI and 17 othersVillage LLM Atlasatlas.village.dev

Margins

Each benchmark scaled 0–100% (its declared range, else the field’s range); Δ is Gemini 4 Carbon minus Claude Opus 5.

ARC-AGI-3 (Semi-Private)General QA30.299.1+68.9
Terminal-Bench-Science 0.1Science28.662.4+33.8
ExploitGym - release-reported percentage, cohort unspecifiedCybersecurity22.154.8+32.7
Bug Hunt BenchCode25.754.6+28.9
Blueprint-Bench 2Multimodal0.30.6+28.2
Terminal-Bench 4.0 (Artificial Analysis)Code34.360.6+26.3
AutomationBench 1.0.6 (Zapier)Agents26.953.0+26.0
Humanity's Last Exam DiamondGeneral QA44.869.1+24.3
AutomationBench-AA 1.0.6Agents56.677.8+21.2
LMArena Agent Arena Net Improvement (Steering Burden)Agents7.217.0+18.4
FrontierSWE V2 Worst@5Code39.156.2+17.1
TxBench Antibody Discovery V1 (100 evaluations)Science48.765.0+16.3
Artificial Analysis Intelligence Index v5.0.1Indexes44.260.0+15.9
BuseyBench SVGMultimodal6.37.9+15.6
FrontierCode 2.0 - Pass RateCode31.446.9+15.6
FrontierCode 2.0 - ScoreCode28.142.6+14.5
LMArena Agent Arena Praise vs ComplaintAgents12.241.0+14.4
CursorBench 4.0Code46.660.4+13.8
APEX-Agents 1.1Professional Work77.090.4+13.4
TxBench Oligo Discovery V1 (113 evaluations)Science42.555.8+13.3
RiemannBenchMath68.080.0+12.0
τ³-Banking (Mercor run, bm25_grep retrieval)Agents44.055.7+11.7
CritPt v2 (agentic, avg@5)Science71.482.3+10.9
LMArena Code Arena WebDevCode16911768+10.1
SWE-sweepCode1.39.6+8.3
GDP.pdf (Artificial Analysis, All-pass)Professional Work21.629.6+8.0
AA-Omniscience v1.1 IndexIndexes37.653.5+7.9
LMArena Agent Arena Bash ReliabilityAgents9.8-5.2−7.5
DAXBench v2Code86.293.6+7.4
LMArena Agent Arena Steering BurdenAgents4.419.1+7.3
APEX-Accounting 1.1Finance & Accounting64.370.7+6.4
LMArena Text Arena EloChat & Writing14901526+6.3
FrontierSWE V2 Best@5Code60.766.7+6.0
Artificial Analysis Progress Index v1Indexes334351+6.0
GDPval-AA v2.1Professional Work16051733+5.6
Medical Long Context Reasoning (MLCR-AA, October 2026 methodology)Health & Medicine13.919.4+5.6
Vals RSI Index v1.1 (four tasks)Agents33.027.7−5.3
MMMU ProMultimodal84.089.2+5.3
LMArena Agent Arena Confirmed SuccessAgents10.218.4+4.1
ARC-AGI-2General QA90.494.2+3.7
TxBench Preclinical Pharmacology V1 Clean RerunsScience63.066.0+3.0
AA-Briefcase v1.1 EloAgents15581611+2.9
AA-LCR v1.1Search82.079.3−2.7
FoodTruck BenchGames7526477218+2.5
OSWorld 2.1 Offline, 500-step Partial ScoreAgents79.281.5+2.3
ARC-AGI-1 Semi-PrivateGeneral QA97.598.5+1.0
Vibe Code Bench v1.1Code88.489.1+0.7
Legal Research BenchLegal55.355.8+0.5
Tax Agent BenchFinance & Accounting76.776.4−0.2
LMArena Agent Arena Tool HallucinationAgents0.90.5−0.2

Claude Opus 5vsGemini 4 Carbonnormalised score marginsData Nov 15, 2026Data: Artificial Analysis, Arena, Vals AI and 17 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…