Atlas

Compare Models

Opus 4.8 vs Gemini 4 Carbon

Gemini 4 Carbon leads by 8.2 capability points (≈95% interval +5.8 to +10.6): a clear lead over Opus 4.8 · P(higher) ≈ 100%

Wins 46 of 47 shared benchmarks (0 lost, 1 tied)

Capability vs. relative cost
Opus 4.8Gemini 4 CarbonClosed weightsOpen weights

Benchmark wins0 – 46 · 47 shared, 1 tied

Comparing matching setups only (5 benchmarks excluded) · suites counted once
← Opus 4.8Gemini 4 Carbon →
Data: Epoch AI, Artificial Analysis, Vals AI and 17 othersatlas.village.dev

Margins

Each benchmark scaled 0–100% (its declared range, else the field’s range); Δ is Gemini 4 Carbon minus Opus 4.8.

ARC-AGI-3 (Semi-Private)Games1.599.1+97.6
HieroglyphBenchPerception & media › Documents & charts14.076.0+62.0
Humanity's Sixth SensePerception & media › Spatial & visual reasoning23.571.1+47.6
Humanity's Last Exam DiamondKnowledge & reasoning › General knowledge23.169.1+46.0
Blueprint-Bench 2Perception & media › Spatial & visual reasoning0.10.6+44.1
Terminal-Bench 4.0 (Artificial Analysis)Software › Software engineering21.762.6+40.9
Bug Hunt BenchSoftware › Software engineering14.354.6+40.3
Terminal-Bench 4.0Software › Software engineering23.263.1+39.9
FrontierMath Tier 4 (v2) - PrivateMath56.195.1+39.0
Mystery Game PuzzlesGames36.073.0+37.0
AutomationBench 1.0.6 (Zapier)Tool & computer use › Customer service & operations16.953.0+36.1
Furniture Assembly BenchmarkPerception & media › Spatial & visual reasoning42.577.5+35.0
RiemannBenchMath47.280.0+32.8
AutomationBench-AA 1.0.6Tool & computer use › Customer service & operations45.677.8+32.2
TxBench Antibody Discovery V1 (100 evaluations)Science › Life sciences33.365.0+31.7
FrontierCode 2.0 - Pass RateSoftware › Software engineering17.146.9+29.8
LMArena Code Arena WebDevSoftware › Frontend & design15551769+28.4
DeepsecBench — Revalidated V1 (233 issues)Software › Cybersecurity6.935.2+28.3
EnterpriseBench: CoreCraftTool & computer use › Customer service & operations52.380.5+28.2
SimpleBenchKnowledge & reasoning › General knowledge64.892.6+27.8
FrontierCode 2.0 - ScoreSoftware › Software engineering15.342.6+27.3
BuseyBench SVGPerception & media › Generation5.27.9+26.5
APEX-Agents 1.1Professional › General knowledge work64.490.4+26.0
Chess Puzzles (task version 2.0)Games52.078.0+26.0
DrugDiscoveryBenchScience › Life sciences46.871.5+24.7
SimpleQA Verified (Epoch anti-abstention)Knowledge & reasoning › General knowledge53.077.7+24.7
ARC-AGI-2Knowledge & reasoning › Abstract reasoning72.194.2+22.1
TxBench Oligo Discovery V1 (113 evaluations)Science › Life sciences37.555.8+18.3
AA-Briefcase v1.1 EloProfessional › General knowledge work13201611+15.9
GBENCH One-shot Coding GScoreGames53.869.0+15.2
HiL-Bench (Human-in-Loop Benchmark)Tool & computer use › General agents35.350.3+15.0
Legal Research BenchProfessional › Legal43.855.8+12.0
GDPval-AA v2.1Professional › General knowledge work14611733+12.0
Tax Agent BenchProfessional › Finance, accounting & tax64.976.4+11.5
Artificial Analysis Progress Index v1Composite indexes318351+11.4
FrontierMath Tiers 1-3 (v2) - PrivateMath80.090.5+10.5
LLM Debate BenchmarkLanguage & writing › Language & wordplay16101676+9.4
TxBench Preclinical Pharmacology V1 Clean RerunsScience › Life sciences57.766.0+8.3
GPQA DiamondScience › Multidisciplinary88.496.5+8.1
Epoch Capabilities Index (ECI)Composite indexes160170+7.9
LMArena Text Arena EloLanguage & writing › General assistance14821526+7.7
GDP.pdf (Artificial Analysis, All-pass)Professional › General knowledge work22.829.6+6.8
Vibe Code Bench v1.1Software › Frontend & design82.789.1+6.3
ARC-AGI-1 Semi-PrivateKnowledge & reasoning › Abstract reasoning92.598.5+6.0
Vals RSI Index v1.1 (four tasks)Software › Data & ML24.027.7+3.8
OTIS Mock AIME 2024-2025Math98.3100.0+1.7
Medical Long Context Reasoning (MLCR-AA, October 2026 methodology)Health & medicine19.419.40.0
Data: Epoch AI, Artificial Analysis, Vals AI and 17 othersatlas.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…