Atlas

Compare Models

Claude Opus 5.5 vs Gemini 4 Hydrogen

Claude Opus 5.5 leads by 1.9 capability points (≈95% interval −0.8 to +4.6): statistically tied · P(higher) ≈ 92%

Wins 43 of 71 shared benchmarks (24 lost, 4 tied)

Capability vs. relative cost
Claude Opus 5.5Gemini 4 HydrogenClosed weightsOpen weights

Benchmark wins43 – 24 · 71 shared, 4 tied

Comparing matching setups only (5 benchmarks excluded) · suites counted once
← Claude Opus 5.5Gemini 4 Hydrogen →
Data: Surge AI, Epoch AI, Arena and 25 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 5.5.

TxBench Oligo Discovery V1 (113 evaluations)Science › Life sciences1.554.3+52.8
TxBench Antibody Discovery V1 (100 evaluations)Science › Life sciences8.357.7+49.3
TxBench Preclinical Pharmacology V1 Clean RerunsScience › Life sciences26.361.3+35.0
Furniture Assembly BenchmarkPerception & media › Spatial & visual reasoning83.357.5−25.8
DAYJOB: Healthcare (October 2026 rescoring)Health & medicine48.324.0−24.3
SWE-Marathon v1.2Software › Software engineering81.358.1−23.1
Humanity's Sixth SensePerception & media › Spatial & visual reasoning44.663.4+18.8
LMArena Code Arena WebDevSoftware › Frontend & design18181703−15.4
sudo L7 (October 2026 rescoring)Software › Software engineering46.131.7−14.4
DAYJOB: Law (November 2026 release)Professional › Legal23.79.7−14.0
Terminal-Bench-Science 0.1Science › Multidisciplinary61.949.5−12.4
GDP.xlsxProfessional › General knowledge work30.340.9+10.6
AA-Briefcase v1.1 EloProfessional › General knowledge work17601569−10.5
FrontierSWE V2 Best@5Software › Software engineering71.561.2−10.3
LMArena Agent Arena Net Improvement (Steering Burden)Tool & computer use › General agents18.613.5−10.0
ComplexConstraints (November 2026 rescoring)Language & writing › Instruction following64.473.7+9.3
LMArena Agent Arena Bash ReliabilityTool & computer use › General agents14.0-4.5−9.3
FrontierCode 2.0 - Pass RateSoftware › Software engineering51.843.2−8.6
Bug Hunt BenchSoftware › Software engineering39.731.1−8.6
FrontierCode 2.0 - ScoreSoftware › Software engineering47.338.9−8.5
HANDBOOK.md Agents v1.1Tool & computer use › Customer service & operations52.344.6−7.7
FrontierMath Tier 4 (v2) - PrivateMath95.087.8−7.2
Chess Puzzles (task version 2.0)Games67.060.0−7.0
FrontierSWE V2 Worst@5Software › Software engineering52.045.3−6.6
LiveBench 2.0 (2026-11-06 release)Composite indexes83.477.0−6.4
GBENCH One-shot Coding GScoreGames73.066.8−6.2
Terminal-Bench 4.0 (Artificial Analysis)Software › Software engineering59.653.5−6.1
Mystery Game PuzzlesGames71.065.0−6.0
DAYJOB: TaxProfessional › Finance, accounting & tax9.23.6−5.6
Professional Reasoning Benchmark - LegalProfessional › Legal47.653.1+5.5
GDP.pdf (November 2026 rescoring)Professional › General knowledge work48.854.2+5.4
LMArena Agent Arena Praise vs ComplaintTool & computer use › General agents38.629.3−4.7
FORTRESS Average Risk Score (ARS)Safety & alignment13.217.5−4.3
Humanity's Last Exam DiamondKnowledge & reasoning › General knowledge54.658.8+4.2
DAYJOB: Finance (October 2026 rescoring)Professional › Finance, accounting & tax42.538.5−4.0
GDP.pdf (Artificial Analysis, All-pass)Professional › General knowledge work28.832.8+4.0
Ramp SWE-Bench v2Software › Software engineering67.871.6+3.8
HiL-Bench (Human-in-Loop Benchmark)Tool & computer use › General agents59.055.3−3.7
CursorBench 4.0Software › Software engineering57.854.3−3.5
SWE-sweepSoftware › Software engineering10.113.5+3.4
CursorBench 4.1Software › Software engineering58.161.4+3.3
SimpleBenchKnowledge & reasoning › General knowledge88.491.7+3.3
FrontierMath Tiers 1-3 (v2) - PrivateMath91.288.1−3.2
SWE-Marathon v1.1Software › Software engineering46.349.4+3.1
LMArena Text Arena EloLanguage & writing › General assistance15071490−3.0
Chartography (November 2026 rescoring)Perception & media › Documents & charts89.687.2−2.4
HLAB-Vals-1.1Professional › Legal9.211.3+2.1
TaxCalcBench TY25Professional › Finance, accounting & tax42.044.0+2.0
SpireClimbGames5.87.7+1.8
LMArena Agent Arena Confirmed SuccessTool & computer use › General agents19.215.7−1.8
GradeBenchProfessional › Education53.551.9−1.6
Epoch Capabilities Index (ECI)Composite indexes169167−1.5
SimpleQA Verified (Epoch anti-abstention)Knowledge & reasoning › General knowledge72.270.9−1.3
BigFinanceBench private holdout (final-answer accuracy)Professional › Finance, accounting & tax48.249.4+1.2
CWE-bench v2 Pass@1Software › Cybersecurity67.168.1+1.0
DualEntry Accounting AI BenchmarkProfessional › Finance, accounting & tax82.283.2+1.0
Gray Swan IPI Attack Success (k=15)Safety & alignment1.00.0+1.0
EBR-bench v4 (card ban, single agent)Games71.470.5−1.0
CritPt v2 (agentic, avg@5)Science › Physical sciences81.480.6−0.9
ARC-AGI-2Knowledge & reasoning › Abstract reasoning93.392.5−0.8
DAYJOB: AccountingProfessional › Finance, accounting & tax34.433.6−0.8
micro1 PharmaBenchProfessional › General knowledge work10.811.4+0.6
Medical Long Context Reasoning (MLCR-AA, October 2026 methodology)Health & medicine27.827.2−0.6
LMArena Agent Arena Steering BurdenTool & computer use › General agents18.017.4−0.3
LMArena Agent Arena Tool HallucinationTool & computer use › Tool calling1.00.7−0.1
BuseyBench SVGPerception & media › Generation6.86.9+0.1
Excel Modeling Benchmark v1.1Professional › Finance, accounting & tax75.875.80.0
Agents' Last Exam v1.1 (ALE v1.1)Professional › General knowledge work43.443.40.0
ARC-AGI-1 Semi-PrivateKnowledge & reasoning › Abstract reasoning98.598.50.0
Dual-Arm Sim (task completion)Tool & computer use › Robotics0.00.00.0
OTIS Mock AIME 2024-2025Math100.0100.00.0
Data: Surge AI, Epoch AI, Arena and 25 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…