Claude Fable 5.2 vs Gemini 4 Hydrogen
Claude Fable 5.2 leads by 4.1 capability points (≈95% interval +1.2 to +6.9): a clear lead over Gemini 4 Hydrogen · P(higher) ≈ 100%
Wins 50 of 68 shared benchmarks (17 lost, 1 tied)
Capability vs. relative cost
Claude Fable 5.2Gemini 4 HydrogenClosed weightsOpen weights
Benchmark wins50 – 17 · 68 shared, 1 tied
Comparing matching setups only (3 benchmarks excluded) · suites counted once
Area← Claude Fable 5.2Gemini 4 Hydrogen →
Data: Epoch AI, Surge AI, Arena and 24 othersatlas.village.dev
Margins
Each benchmark scaled 0–100% (its declared range, else the field’s range); Δ is Gemini 4 Hydrogen minus Claude Fable 5.2.
| Furniture Assembly Benchmark | Perception & media › Spatial & visual reasoning | 90.0 | 57.5 | −32.5 |
| SWE-Marathon v1.2 | Software › Software engineering | 86.8 | 58.1 | −28.7 |
| LMArena Agent Arena Net Improvement (Steering Burden) | Tool & computer use › General agents | 27.4 | 13.5 | −27.1 |
| Mystery Game Puzzles | Games | 90.0 | 65.0 | −25.0 |
| Bug Hunt Bench | Software › Software engineering | 55.2 | 31.1 | −24.1 |
| SimpleQA Verified (Epoch anti-abstention) | Knowledge & reasoning › General knowledge | 94.0 | 70.9 | −23.1 |
| HiL-Bench (Human-in-Loop Benchmark) | Tool & computer use › General agents | 77.3 | 55.3 | −22.0 |
| Humanity's Last Exam Diamond | Knowledge & reasoning › General knowledge | 80.2 | 58.8 | −21.4 |
| DAYJOB: Healthcare (October 2026 rescoring) | Health & medicine | 45.0 | 24.0 | −21.0 |
| sudo L7 (October 2026 rescoring) | Software › Software engineering | 52.2 | 31.7 | −20.5 |
| EBR-bench v4 (card ban, single agent) | Games | 90.5 | 70.5 | −20.0 |
| LMArena Agent Arena Praise vs Complaint | Tool & computer use › General agents | 68.7 | 29.3 | −19.7 |
| Chess Puzzles (task version 2.0) | Games | 79.0 | 60.0 | −19.0 |
| TaxCalcBench TY25 | Professional › Finance, accounting & tax | 62.0 | 44.0 | −18.0 |
| FrontierSWE V2 | Software › Software engineering | 69.9 | 53.1 | −16.8 |
| DAYJOB: Law (November 2026 release) | Professional › Legal | 26.5 | 9.7 | −16.8 |
| Medical Long Context Reasoning (MLCR-AA, October 2026 methodology) | Health & medicine | 43.9 | 27.2 | −16.7 |
| LMArena Code Arena WebDev | Software › Frontend & design | 1824 | 1703 | −16.1 |
| FrontierMath Tier 4 (v2) - Private | Math | 100.0 | 87.8 | −12.2 |
| DAYJOB: Tax | Professional › Finance, accounting & tax | 15.6 | 3.6 | −12.0 |
| DAYJOB: Accounting | Professional › Finance, accounting & tax | 45.2 | 33.6 | −11.6 |
| FrontierCode 2.0 - Pass Rate | Software › Software engineering | 54.7 | 43.2 | −11.5 |
| Tax Agent Bench | Professional › Finance, accounting & tax | 67.0 | 78.4 | +11.4 |
| FrontierCode 2.0 - Score | Software › Software engineering | 50.2 | 38.9 | −11.3 |
| AA-Briefcase v1.1 Elo | Professional › General knowledge work | 1737 | 1569 | −9.2 |
| ComplexConstraints (November 2026 rescoring) | Language & writing › Instruction following | 64.5 | 73.7 | +9.2 |
| LMArena Agent Arena Bash Reliability | Tool & computer use › General agents | 13.1 | -4.5 | −8.8 |
| GBENCH One-shot Coding GScore | Games | 74.9 | 66.8 | −8.1 |
| GDP.xlsx | Professional › General knowledge work | 32.9 | 40.9 | +8.0 |
| LiveBench 2.0 (2026-11-06 release) | Composite indexes | 84.9 | 77.0 | −7.9 |
| HANDBOOK.md Agents v1.1 | Tool & computer use › Customer service & operations | 52.3 | 44.6 | −7.7 |
| SpireClimb | Games | 15.3 | 7.7 | −7.7 |
| Dual-Arm Sim (task completion) | Tool & computer use › Robotics | 7.5 | 0.0 | −7.5 |
| FrontierMath Tiers 1-3 (v2) - Private | Math | 95.1 | 88.1 | −7.0 |
| CWE-bench v2 Pass@1 | Software › Cybersecurity | 75.0 | 68.1 | −6.9 |
| AutomationBench 1.0.6 (Zapier) | Tool & computer use › Customer service & operations | 44.3 | 51.1 | +6.8 |
| Chartography | Perception & media › Documents & charts | 73.4 | 66.6 | −6.8 |
| HLAB-Vals-1.1 | Professional › Legal | 5.0 | 11.3 | +6.3 |
| KradleBench | Games | 37.8 | 43.9 | +6.1 |
| Vibe Code Bench v1.1 | Software › Frontend & design | 93.8 | 87.8 | −6.0 |
| GradeBench | Professional › Education | 57.3 | 51.9 | −5.4 |
| Epoch Capabilities Index (ECI) | Composite indexes | 173 | 167 | −5.3 |
| Professional Reasoning Benchmark - Finance | Professional › Finance, accounting & tax | 47.1 | 52.3 | +5.2 |
| DAYJOB: Finance (October 2026 rescoring) | Professional › Finance, accounting & tax | 43.7 | 38.5 | −5.2 |
| Finance Agent v2.1 | Professional › Finance, accounting & tax | 69.5 | 65.1 | −4.4 |
| CritPt v2 (agentic, avg@5) | Science › Physical sciences | 84.9 | 80.6 | −4.3 |
| Professional Reasoning Benchmark - Legal | Professional › Legal | 49.0 | 53.1 | +4.1 |
| Gray Swan IPI Attack Success (k=15) | Safety & alignment | 4.1 | 0.0 | +4.1 |
| Humanity's Sixth Sense | Perception & media › Spatial & visual reasoning | 59.5 | 63.4 | +3.9 |
| LMArena Agent Arena Steering Burden | Tool & computer use › General agents | 24.5 | 17.4 | −3.5 |
| Agents' Last Exam v1.1 (ALE v1.1) | Professional › General knowledge work | 46.9 | 43.4 | −3.5 |
| LMArena Agent Arena Confirmed Success | Tool & computer use › General agents | 22.5 | 15.7 | −3.4 |
| SimpleBench | Knowledge & reasoning › General knowledge | 94.9 | 91.7 | −3.2 |
| BuseyBench SVG | Perception & media › Generation | 7.2 | 6.9 | −3.1 |
| GPQA Diamond | Science › Multidisciplinary | 93.9 | 96.8 | +2.9 |
| Vals Finance Index | Professional › Finance, accounting & tax | 63.6 | 66.1 | +2.6 |
| GDP.pdf (November 2026 rescoring) | Professional › General knowledge work | 51.8 | 54.2 | +2.4 |
| DualEntry Accounting AI Benchmark | Professional › Finance, accounting & tax | 81.2 | 83.2 | +2.0 |
| FORTRESS Average Risk Score (ARS) | Safety & alignment | 15.6 | 17.5 | −1.9 |
| Ramp SWE-Bench v2 | Software › Software engineering | 73.4 | 71.6 | −1.8 |
| Excel Modeling Benchmark v1.1 | Professional › Finance, accounting & tax | 77.4 | 75.8 | −1.6 |
| ARC-AGI-2 | Knowledge & reasoning › Abstract reasoning | 93.8 | 92.5 | −1.2 |
| micro1 PharmaBench | Professional › General knowledge work | 12.2 | 11.4 | −0.8 |
| OTIS Mock AIME 2024-2025 | Math | 99.7 | 100.0 | +0.3 |
| AutomationBench-AA 1.0.6 | Tool & computer use › Customer service & operations | 74.2 | 74.4 | +0.3 |
| LMArena Agent Arena Tool Hallucination | Tool & computer use › Tool calling | 1.2 | 0.7 | −0.2 |
| MultiChallenge | Language & writing › Conversation | 80.4 | 80.6 | +0.2 |
| ARC-AGI-1 Semi-Private | Knowledge & reasoning › Abstract reasoning | 98.5 | 98.5 | 0.0 |
Data: Epoch AI, Surge AI, Arena and 24 othersatlas.village.dev