Gemini 4 Carbon vs GPT-5.5
Gemini 4 Carbon leads by 7.9 capability points (≈95% interval +5.5 to +10.3): a clear lead over GPT-5.5 · P(higher) ≈ 100%
Wins 43 of 46 shared benchmarks (2 lost, 1 tied)
Capability vs. relative cost
Gemini 4 CarbonGPT-5.5Closed weightsOpen weights
Benchmark wins43 – 2 · 46 shared, 1 tied
Comparing matching setups only (14 benchmarks excluded) · suites counted once
Area← Gemini 4 CarbonGPT-5.5 →
Data: Epoch AI, Artificial Analysis, ARC Prize Foundation and 16 othersatlas.village.dev
Margins
Each benchmark scaled 0–100% (its declared range, else the field’s range); Δ is GPT-5.5 minus Gemini 4 Carbon.
| ARC-AGI-3 (Semi-Private) | Games | 99.1 | 0.4 | −98.7 |
| Terminal-Bench 4.0 (Artificial Analysis) | Software › Software engineering | 60.6 | 5.1 | −55.6 |
| HieroglyphBench | Perception & media › Documents & charts | 76.0 | 22.8 | −53.2 |
| Chartography | Perception & media › Documents & charts | 72.5 | 31.0 | −41.5 |
| Humanity's Last Exam Diamond | Knowledge & reasoning › General knowledge | 69.1 | 27.9 | −41.2 |
| AutomationBench 1.0.6 (Zapier) | Tool & computer use › Customer service & operations | 53.0 | 16.6 | −36.4 |
| Furniture Assembly Benchmark | Perception & media › Spatial & visual reasoning | 77.5 | 44.2 | −33.3 |
| TxBench Antibody Discovery V1 (100 evaluations) | Science › Life sciences | 65.0 | 33.0 | −32.0 |
| AutomationBench-AA 1.0.6 | Tool & computer use › Customer service & operations | 77.8 | 47.3 | −30.5 |
| EnterpriseBench: CoreCraft | Tool & computer use › Customer service & operations | 80.5 | 51.3 | −29.2 |
| AA-Briefcase v1.1 Elo | Professional › General knowledge work | 1611 | 1087 | −28.7 |
| RiemannBench | Math | 80.0 | 55.2 | −24.8 |
| LLM Debate Benchmark | Language & writing › Language & wordplay | 1676 | 1507 | −24.2 |
| SimpleBench | Knowledge & reasoning › General knowledge | 92.6 | 69.0 | −23.6 |
| FrontierMath Tier 4 (v2) - Private | Math | 95.1 | 72.5 | −22.6 |
| Blueprint-Bench 2 | Perception & media › Spatial & visual reasoning | 0.6 | 0.4 | −22.4 |
| TxBench Oligo Discovery V1 (113 evaluations) | Science › Life sciences | 55.8 | 33.9 | −21.8 |
| APEX-Agents 1.1 | Professional › General knowledge work | 90.4 | 70.3 | −20.1 |
| DrugDiscoveryBench | Science › Life sciences | 71.5 | 51.6 | −19.9 |
| Vibe Code Bench v1.1 | Software › Frontend & design | 89.1 | 69.8 | −19.2 |
| BuseyBench SVG | Perception & media › Generation | 7.9 | 6.0 | −18.6 |
| GDPval-AA v2.1 | Professional › General knowledge work | 1733 | 1335 | −17.5 |
| Mystery Game Puzzles | Games | 73.0 | 56.0 | −17.0 |
| Tax Agent Bench | Professional › Finance, accounting & tax | 76.4 | 60.5 | −16.0 |
| Legal Research Bench | Professional › Legal | 55.8 | 40.4 | −15.4 |
| Chess Puzzles (task version 2.0) | Games | 78.0 | 63.0 | −15.0 |
| SimpleQA Verified (Epoch anti-abstention) | Knowledge & reasoning › General knowledge | 77.7 | 63.0 | −14.7 |
| APEX-Accounting 1.1 | Professional › Finance, accounting & tax | 70.7 | 56.3 | −14.4 |
| DeepsecBench — Revalidated V1 (233 issues) | Software › Cybersecurity | 35.2 | 21.0 | −14.2 |
| Artificial Analysis Progress Index v1 | Composite indexes | 351 | 320 | −10.8 |
| TxBench Preclinical Pharmacology V1 Clean Reruns | Science › Life sciences | 66.0 | 55.3 | −10.7 |
| HiL-Bench (Human-in-Loop Benchmark) | Tool & computer use › General agents | 50.3 | 39.7 | −10.7 |
| Medical Long Context Reasoning (MLCR-AA, October 2026 methodology) | Health & medicine | 19.4 | 28.9 | +9.4 |
| ARC-AGI-2 | Knowledge & reasoning › Abstract reasoning | 94.2 | 85.0 | −9.2 |
| MMMU Pro | Knowledge & reasoning › General knowledge | 89.2 | 81.2 | −8.1 |
| LMArena Text Arena Elo | Language & writing › General assistance | 1526 | 1481 | −7.8 |
| Epoch Capabilities Index (ECI) | Composite indexes | 170 | 160 | −7.6 |
| GBENCH One-shot Coding GScore | Games | 69.0 | 61.6 | −7.4 |
| GDP.pdf (Artificial Analysis, All-pass) | Professional › General knowledge work | 29.6 | 22.6 | −7.0 |
| GPQA Diamond | Science › Multidisciplinary | 96.5 | 90.7 | −5.9 |
| DAXBench v2 | Software › Data & ML | 93.6 | 88.3 | −5.3 |
| FrontierMath Tiers 1-3 (v2) - Private | Math | 90.5 | 85.3 | −5.3 |
| AA-LCR v1.1 | Knowledge & reasoning › Reading & long context | 79.3 | 84.3 | +5.0 |
| ARC-AGI-1 Semi-Private | Knowledge & reasoning › Abstract reasoning | 98.5 | 95.0 | −3.5 |
| FrontierMath Erdős | Math | 2.9 | 0.0 | −2.9 |
| OTIS Mock AIME 2024-2025 | Math | 100.0 | 100.0 | 0.0 |
Data: Epoch AI, Artificial Analysis, ARC Prize Foundation and 16 othersatlas.village.dev