Atlas

Compare Models

Claude Opus 5 vs Gemini 4 Hydrogen

Gemini 4 Hydrogen leads by 1.8 capability points (≈95% interval −0.7 to +4.3): statistically tied · P(higher) ≈ 92%

Wins 43 of 55 shared benchmarks (12 lost)

Capability vs. relative cost
Claude Opus 5Gemini 4 HydrogenClosed weightsOpen weights

Benchmark wins12 – 43 · 55 shared

Comparing matching setups only (3 benchmarks excluded) · suites counted once
← Claude Opus 5Gemini 4 Hydrogen →
Data: Artificial Analysis, Epoch AI, Arena and 19 othersatlas.village.dev

Margins

Each benchmark scaled 0–100% (its declared range, else the field’s range); Δ is Gemini 4 Hydrogen minus Claude Opus 5.

Humanity's Sixth SensePerception & media › Spatial & visual reasoning30.563.4+32.9
EBR-bench v4 (card ban, single agent)Games45.770.5+24.8
AutomationBench 1.0.6 (Zapier)Tool & computer use › Customer service & operations26.951.1+24.2
TaxCalcBench TY25Professional › Finance, accounting & tax20.044.0+24.0
Terminal-Bench-Science 0.1Science › Multidisciplinary28.651.4+22.8
AutomationBench-AA 1.0.6Tool & computer use › Customer service & operations56.674.4+17.9
Chess Puzzles (task version 2.0)Games43.060.0+17.0
DAYJOB: Finance (October 2026 rescoring)Professional › Finance, accounting & tax23.838.5+14.7
FrontierMath Tier 4 (v2) - PrivateMath73.287.8+14.6
LMArena Agent Arena Net Improvement (Steering Burden)Tool & computer use › General agents6.313.5+14.1
Humanity's Last Exam DiamondKnowledge & reasoning › General knowledge44.858.8+14.0
Medical Long Context Reasoning (MLCR-AA, October 2026 methodology)Health & medicine13.927.2+13.3
FrontierCode 2.0 - Pass RateSoftware › Software engineering31.443.2+11.8
TxBench Oligo Discovery V1 (113 evaluations)Science › Life sciences42.554.3+11.8
GDP.pdf (Artificial Analysis, All-pass)Professional › General knowledge work21.632.8+11.2
SimpleBenchKnowledge & reasoning › General knowledge80.691.7+11.1
FrontierCode 2.0 - ScoreSoftware › Software engineering28.138.9+10.7
LMArena Agent Arena Praise vs ComplaintTool & computer use › General agents10.329.3+9.5
CritPt v2 (agentic, avg@5)Science › Physical sciences71.480.6+9.1
TxBench Antibody Discovery V1 (100 evaluations)Science › Life sciences48.757.7+9.0
GPQA DiamondScience › Multidisciplinary87.996.8+8.9
Terminal-Bench 4.0 (Artificial Analysis)Software › Software engineering46.054.0+8.1
CursorBench 4.0Software › Software engineering46.654.3+7.7
LMArena Agent Arena Steering BurdenTool & computer use › General agents2.417.4+7.5
LMArena Agent Arena Bash ReliabilityTool & computer use › General agents8.7-4.5−6.6
Mystery Game PuzzlesGames59.065.0+6.0
BuseyBench SVGPerception & media › Generation6.36.9+5.6
Bug Hunt BenchSoftware › Software engineering25.731.1+5.4
AA-Omniscience v1.1 IndexKnowledge & reasoning › General knowledge37.646.3+4.3
Furniture Assembly BenchmarkPerception & media › Spatial & visual reasoning60.857.5−3.3
AA-Briefcase v1.1 EloProfessional › General knowledge work16291569−3.3
Artificial Analysis Intelligence Index v5.0.1Composite indexes44.241.2−2.9
LMArena Agent Arena Confirmed SuccessTool & computer use › General agents10.115.7+2.8
Chartography (November 2026 rescoring)Perception & media › Documents & charts84.487.2+2.8
FrontierMath Tiers 1-3 (v2) - PrivateMath85.688.1+2.5
ARC-AGI-2Knowledge & reasoning › Abstract reasoning90.492.5+2.1
DAYJOB: Healthcare (October 2026 rescoring)Health & medicine22.024.0+2.0
Tax Agent BenchProfessional › Finance, accounting & tax76.778.4+1.8
micro1 PharmaBenchProfessional › General knowledge work9.711.4+1.7
Epoch Capabilities Index (ECI)Composite indexes165167+1.7
TxBench Preclinical Pharmacology V1 Clean RerunsScience › Life sciences63.061.3−1.7
HiL-Bench (Human-in-Loop Benchmark)Tool & computer use › General agents57.055.3−1.7
OSWorld 2.1 Offline, 500-step Partial ScoreTool & computer use › Computer & browser use79.277.6−1.6
LMArena Code Arena WebDevSoftware › Frontend & design16911703+1.5
Legal Research BenchProfessional › Legal55.353.8−1.4
OTIS Mock AIME 2024-2025Math98.9100.0+1.1
FrontierSWE V2Software › Software engineering52.053.1+1.1
ARC-AGI-1 Semi-PrivateKnowledge & reasoning › Abstract reasoning97.598.5+1.0
Vibe Code Bench v1.1Software › Frontend & design88.487.8−0.6
SWE-Marathon v1.1Software › Software engineering48.849.4+0.6
FORTRESS Average Risk Score (ARS)Safety & alignment16.817.5−0.6
GBENCH One-shot Coding GScoreGames67.166.8−0.3
BigFinanceBench private holdout (final-answer accuracy)Professional › Finance, accounting & tax49.249.4+0.2
LMArena Agent Arena Tool HallucinationTool & computer use › Tool calling1.00.7−0.2
LMArena Text Arena EloLanguage & writing › General assistance149014900.0
Data: Artificial Analysis, Epoch AI, Arena and 19 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…