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
Area← 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) | Games | 1.5 | 99.1 | +97.6 |
| HieroglyphBench | Perception & media › Documents & charts | 14.0 | 76.0 | +62.0 |
| Humanity's Sixth Sense | Perception & media › Spatial & visual reasoning | 23.5 | 71.1 | +47.6 |
| Humanity's Last Exam Diamond | Knowledge & reasoning › General knowledge | 23.1 | 69.1 | +46.0 |
| Blueprint-Bench 2 | Perception & media › Spatial & visual reasoning | 0.1 | 0.6 | +44.1 |
| Terminal-Bench 4.0 (Artificial Analysis) | Software › Software engineering | 21.7 | 62.6 | +40.9 |
| Bug Hunt Bench | Software › Software engineering | 14.3 | 54.6 | +40.3 |
| Terminal-Bench 4.0 | Software › Software engineering | 23.2 | 63.1 | +39.9 |
| FrontierMath Tier 4 (v2) - Private | Math | 56.1 | 95.1 | +39.0 |
| Mystery Game Puzzles | Games | 36.0 | 73.0 | +37.0 |
| AutomationBench 1.0.6 (Zapier) | Tool & computer use › Customer service & operations | 16.9 | 53.0 | +36.1 |
| Furniture Assembly Benchmark | Perception & media › Spatial & visual reasoning | 42.5 | 77.5 | +35.0 |
| RiemannBench | Math | 47.2 | 80.0 | +32.8 |
| AutomationBench-AA 1.0.6 | Tool & computer use › Customer service & operations | 45.6 | 77.8 | +32.2 |
| TxBench Antibody Discovery V1 (100 evaluations) | Science › Life sciences | 33.3 | 65.0 | +31.7 |
| FrontierCode 2.0 - Pass Rate | Software › Software engineering | 17.1 | 46.9 | +29.8 |
| LMArena Code Arena WebDev | Software › Frontend & design | 1555 | 1769 | +28.4 |
| DeepsecBench — Revalidated V1 (233 issues) | Software › Cybersecurity | 6.9 | 35.2 | +28.3 |
| EnterpriseBench: CoreCraft | Tool & computer use › Customer service & operations | 52.3 | 80.5 | +28.2 |
| SimpleBench | Knowledge & reasoning › General knowledge | 64.8 | 92.6 | +27.8 |
| FrontierCode 2.0 - Score | Software › Software engineering | 15.3 | 42.6 | +27.3 |
| BuseyBench SVG | Perception & media › Generation | 5.2 | 7.9 | +26.5 |
| APEX-Agents 1.1 | Professional › General knowledge work | 64.4 | 90.4 | +26.0 |
| Chess Puzzles (task version 2.0) | Games | 52.0 | 78.0 | +26.0 |
| DrugDiscoveryBench | Science › Life sciences | 46.8 | 71.5 | +24.7 |
| SimpleQA Verified (Epoch anti-abstention) | Knowledge & reasoning › General knowledge | 53.0 | 77.7 | +24.7 |
| ARC-AGI-2 | Knowledge & reasoning › Abstract reasoning | 72.1 | 94.2 | +22.1 |
| TxBench Oligo Discovery V1 (113 evaluations) | Science › Life sciences | 37.5 | 55.8 | +18.3 |
| AA-Briefcase v1.1 Elo | Professional › General knowledge work | 1320 | 1611 | +15.9 |
| GBENCH One-shot Coding GScore | Games | 53.8 | 69.0 | +15.2 |
| HiL-Bench (Human-in-Loop Benchmark) | Tool & computer use › General agents | 35.3 | 50.3 | +15.0 |
| Legal Research Bench | Professional › Legal | 43.8 | 55.8 | +12.0 |
| GDPval-AA v2.1 | Professional › General knowledge work | 1461 | 1733 | +12.0 |
| Tax Agent Bench | Professional › Finance, accounting & tax | 64.9 | 76.4 | +11.5 |
| Artificial Analysis Progress Index v1 | Composite indexes | 318 | 351 | +11.4 |
| FrontierMath Tiers 1-3 (v2) - Private | Math | 80.0 | 90.5 | +10.5 |
| LLM Debate Benchmark | Language & writing › Language & wordplay | 1610 | 1676 | +9.4 |
| TxBench Preclinical Pharmacology V1 Clean Reruns | Science › Life sciences | 57.7 | 66.0 | +8.3 |
| GPQA Diamond | Science › Multidisciplinary | 88.4 | 96.5 | +8.1 |
| Epoch Capabilities Index (ECI) | Composite indexes | 160 | 170 | +7.9 |
| LMArena Text Arena Elo | Language & writing › General assistance | 1482 | 1526 | +7.7 |
| GDP.pdf (Artificial Analysis, All-pass) | Professional › General knowledge work | 22.8 | 29.6 | +6.8 |
| Vibe Code Bench v1.1 | Software › Frontend & design | 82.7 | 89.1 | +6.3 |
| ARC-AGI-1 Semi-Private | Knowledge & reasoning › Abstract reasoning | 92.5 | 98.5 | +6.0 |
| Vals RSI Index v1.1 (four tasks) | Software › Data & ML | 24.0 | 27.7 | +3.8 |
| OTIS Mock AIME 2024-2025 | Math | 98.3 | 100.0 | +1.7 |
| Medical Long Context Reasoning (MLCR-AA, October 2026 methodology) | Health & medicine | 19.4 | 19.4 | 0.0 |
Data: Epoch AI, Artificial Analysis, Vals AI and 17 othersatlas.village.dev