Gemini 4 Hydrogen vs GPT-5.6 Sol
TryTop 2 now
Gemini 4 Hydrogen leads by 3.1 capability points (≈95% interval +0.7 to +5.5): a clear lead over GPT-5.6 Sol · P(higher) ≈ 99%
Wins 43 of 50 shared benchmarks (4 lost, 3 tied)
Capability vs. relative cost
Gemini 4 HydrogenGPT-5.6 SolClosed weightsOpen weights
Benchmark wins43 – 4 · 50 shared, 3 tied
Comparing matching setups only (13 benchmarks excluded) · suites counted once
Area← Gemini 4 HydrogenGPT-5.6 Sol →
Data: Epoch AI, Artificial Analysis, Arena and 15 othersatlas.village.dev
Margins
Each benchmark scaled 0–100% (its declared range, else the field’s range); Δ is GPT-5.6 Sol minus Gemini 4 Hydrogen.
| TxBench-Antibody-Development V1 (60 evaluations) | Science › Life sciences | 60.6 | 18.9 | −41.7 |
| Humanity's Sixth Sense | Perception & media › Spatial & visual reasoning | 63.4 | 30.0 | −33.4 |
| Terminal-Bench-Science 0.1 | Science › Multidisciplinary | 51.4 | 22.4 | −29.0 |
| Terminal-Bench 4.0 (Artificial Analysis) | Software › Software engineering | 53.5 | 24.7 | −28.8 |
| Humanity's Last Exam Diamond | Knowledge & reasoning › General knowledge | 58.8 | 30.4 | −28.4 |
| Gray Swan IPI Attack Success (k=15) | Safety & alignment | 0.0 | 27.0 | −27.0 |
| DAYJOB: Finance (October 2026 rescoring) | Professional › Finance, accounting & tax | 38.5 | 11.8 | −26.7 |
| EBR-bench v4 (card ban, single agent) | Games | 70.5 | 44.8 | −25.7 |
| HiL-Bench (Human-in-Loop Benchmark) | Tool & computer use › General agents | 55.3 | 32.3 | −23.0 |
| AutomationBench 1.0.6 (Zapier) | Tool & computer use › Customer service & operations | 51.1 | 28.8 | −22.4 |
| Chartography | Perception & media › Documents & charts | 66.6 | 45.0 | −21.6 |
| FrontierSWE V2 | Software › Software engineering | 53.1 | 32.2 | −20.9 |
| CursorBench 4.1 | Software › Software engineering | 61.4 | 40.9 | −20.5 |
| TxBench Antibody Discovery V1 (100 evaluations) | Science › Life sciences | 57.7 | 37.3 | −20.3 |
| DAYJOB: Healthcare (October 2026 rescoring) | Health & medicine | 24.0 | 4.4 | −19.6 |
| LMArena Agent Arena Net Improvement (Steering Burden) | Tool & computer use › General agents | 13.5 | 3.7 | −19.2 |
| TxBench Oligo Discovery V1 (113 evaluations) | Science › Life sciences | 54.3 | 37.2 | −17.1 |
| Terminal-Bench 4.0 | Software › Software engineering | 54.0 | 37.9 | −16.2 |
| AutomationBench-AA 1.0.6 | Tool & computer use › Customer service & operations | 74.4 | 60.1 | −14.4 |
| LMArena Agent Arena Praise vs Complaint | Tool & computer use › General agents | 29.3 | 3.8 | −12.7 |
| DrugDiscoveryBench | Science › Life sciences | 69.1 | 56.5 | −12.6 |
| CursorBench 4.0 | Software › Software engineering | 54.3 | 41.7 | −12.6 |
| TaxCalcBench TY25 | Professional › Finance, accounting & tax | 44.0 | 32.0 | −12.0 |
| Tax Agent Bench | Professional › Finance, accounting & tax | 78.4 | 68.0 | −10.5 |
| LMArena Agent Arena Bash Reliability | Tool & computer use › General agents | -4.5 | 10.5 | +7.5 |
| BuseyBench SVG | Perception & media › Generation | 6.9 | 6.1 | −7.5 |
| AA-Briefcase v1.1 Elo | Professional › General knowledge work | 1569 | 1436 | −7.3 |
| Vibe Code Bench v1.1 | Software › Frontend & design | 87.8 | 80.5 | −7.3 |
| Mystery Game Puzzles | Games | 65.0 | 58.0 | −7.0 |
| GPQA Diamond | Science › Multidisciplinary | 96.8 | 89.9 | −6.9 |
| LMArena Agent Arena Steering Burden | Tool & computer use › General agents | 17.4 | 3.8 | −6.8 |
| LMArena Agent Arena Confirmed Success | Tool & computer use › General agents | 15.7 | 2.5 | −6.6 |
| Legal Research Bench | Professional › Legal | 53.8 | 48.1 | −5.8 |
| BigFinanceBench private holdout (final-answer accuracy) | Professional › Finance, accounting & tax | 49.4 | 44.2 | −5.2 |
| GDP.pdf (Artificial Analysis, All-pass) | Professional › General knowledge work | 32.8 | 27.8 | −5.0 |
| FrontierMath Tier 4 (v2) - Private | Math | 87.8 | 82.9 | −4.9 |
| CritPt v2 (agentic, avg@5) | Science › Physical sciences | 80.6 | 77.1 | −3.4 |
| Epoch Capabilities Index (ECI) | Composite indexes | 167 | 163 | −3.0 |
| Medical Long Context Reasoning (MLCR-AA, October 2026 methodology) | Health & medicine | 27.2 | 30.0 | +2.8 |
| FORTRESS Average Risk Score (ARS) | Safety & alignment | 17.5 | 20.1 | −2.7 |
| FrontierMath Tiers 1-3 (v2) - Private | Math | 88.1 | 89.1 | +1.1 |
| ARC-AGI-1 Semi-Private | Knowledge & reasoning › Abstract reasoning | 98.5 | 97.5 | −1.0 |
| Chess Puzzles (task version 2.0) | Games | 60.0 | 59.0 | −1.0 |
| Furniture Assembly Benchmark | Perception & media › Spatial & visual reasoning | 57.5 | 56.7 | −0.8 |
| LMArena Text Arena Elo | Language & writing › General assistance | 1490 | 1485 | −0.8 |
| GBENCH One-shot Coding GScore | Games | 66.8 | 66.4 | −0.4 |
| LMArena Agent Arena Tool Hallucination | Tool & computer use › Tool calling | 0.7 | 1.2 | +0.2 |
| ARC-AGI-2 | Knowledge & reasoning › Abstract reasoning | 92.5 | 92.5 | 0.0 |
| OTIS Mock AIME 2024-2025 | Math | 100.0 | 100.0 | 0.0 |
| TxBench Preclinical Pharmacology V1 Clean Reruns | Science › Life sciences | 61.3 | 61.3 | 0.0 |
Data: Epoch AI, Artificial Analysis, Arena and 15 othersatlas.village.dev