Atlas

Compare Models

Claude Opus 5.5 vs Gemini 4 Carbon

Gemini 4 Carbon leads by 0.2 capability points (≈95% interval −2.6 to +3.0): statistically tied · P(higher) ≈ 56%

Wins 34 of 50 shared benchmarks (15 lost, 1 tied)

Capability vs. relative cost
Claude Opus 5.5Gemini 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 5.5vsGemini 4 Carboncapability vs. relative costData Nov 15, 2026Village LLM Atlasatlas.village.dev

Benchmark wins15 – 34 · 50 shared, 1 tied

Comparing matching setups only (8 benchmarks excluded) · suites counted once
← Claude Opus 5.5Gemini 4 Carbon →

Claude Opus 5.5vsGemini 4 Carbonwins by categoryData Nov 15, 2026Data: Artificial Analysis, Arena, MathArena and 15 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 5.5.

TxBench Antibody Discovery V1 (100 evaluations)Science8.365.0+56.7
TxBench Oligo Discovery V1 (113 evaluations)Science1.555.8+54.3
TxBench Preclinical Pharmacology V1 Clean RerunsScience26.366.0+39.7
Terminal-Bench-Science 0.1Science43.362.4+19.0
ArXivLean 09/2026Math47.262.3+15.1
Bug Hunt BenchCode39.754.6+14.9
Humanity's Last Exam DiamondGeneral QA54.669.1+14.5
ArXivMath 09/2026 (Native Harnesses)Math82.196.2+14.2
ArXivMath 10/2026 (Native Harnesses)Math79.893.6+13.8
Harvey LAB (Mercor run, undisclosed ~213-task cohort)Legal10.823.0+12.2
BrokenArXiv 08/2026 (Native Harnesses)Math77.488.1+10.7
RiemannBenchMath69.680.0+10.4
LMArena Agent Arena Bash ReliabilityAgents15.1-5.2−10.2
BuseyBench SVGMultimodal6.87.9+10.1
Vals RSI Index v1.1 (four tasks)Agents37.327.7−9.6
APEX-Agents 1.1Professional Work81.390.4+9.1
OpenWorldEngineer (Mean@8)Code41.632.8−8.8
Medical Long Context Reasoning (MLCR-AA, October 2026 methodology)Health & Medicine27.819.4−8.3
HLAB-Vals-1.1Legal9.217.5+8.3
Terminal-Bench 4.0 (Artificial Analysis)Code52.560.6+8.1
Blueprint-Bench 2Multimodal0.50.6+7.4
ArXivMath 08/2026 (Native Harnesses)Math86.093.0+7.0
CritPt v2 (agentic, avg@5)Science76.082.3+6.3
LMArena Code Arena WebDevCode18141768−6.0
BrokenArXiv 09/2026 (Native Harnesses)Math76.782.4+5.7
AA-LCR v1.1Search84.779.3−5.3
APEX-Accounting 1.1Finance & Accounting65.570.7+5.2
BrokenArXiv 10/2026 (Native Harnesses)Math77.882.7+4.9
FrontierCode 2.0 - Pass RateCode51.846.9−4.9
FrontierSWE V2 Best@5Code71.566.7−4.8
LMArena Agent Arena Net Improvement (Steering Burden)Agents19.517.0−4.8
FrontierCode 2.0 - ScoreCode47.342.6−4.7
FrontierSWE V2 Worst@5Code52.056.2+4.2
CWE-bench v1 Programmatic Pass@1Cybersecurity67.071.0+4.0
AA-Briefcase v1.1 EloAgents16821611−3.9
LMArena Text Arena EloChat & Writing15061526+3.5
MMMU ProMultimodal86.689.2+2.7
CursorBench 4.0Code57.860.4+2.6
DAXBench v2Code95.793.6−2.1
GDPxProfessional Work12.010.9−1.2
LMArena Agent Arena Confirmed SuccessAgents20.718.4−1.2
GDPval-AA v2.1Professional Work17091733+1.1
ARC-AGI-2General QA93.394.2+0.8
GDP.pdf (Artificial Analysis, All-pass)Professional Work28.829.6+0.8
τ³-Banking (Mercor run, bm25_grep retrieval)Agents55.055.7+0.7
Excel Modeling Benchmark v1.1Finance & Accounting75.876.5+0.6
LMArena Agent Arena Praise vs ComplaintAgents40.341.0+0.4
LMArena Agent Arena Steering BurdenAgents18.619.1+0.2
LMArena Agent Arena Tool HallucinationAgents0.70.5−0.1
ARC-AGI-1 Semi-PrivateGeneral QA98.598.50.0

Claude Opus 5.5vsGemini 4 Carbonnormalised score marginsData Nov 15, 2026Data: Artificial Analysis, Arena, MathArena and 15 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…