Atlas

Compare Models

Gemini 3.8 Flash vs Gemini 4 Hydrogen

Gemini 4 Hydrogen leads by 5.0 capability points (≈95% interval +2.6 to +7.5): a clear lead over Gemini 3.8 Flash · P(higher) ≈ 100%

Wins 68 of 75 shared benchmarks (4 lost, 3 tied)

Capability vs. relative cost
Gemini 3.8 FlashGemini 4 HydrogenClosed weightsOpen weights

Benchmark wins4 – 68 · 75 shared, 3 tied

Comparing matching setups only (1 benchmark excluded) · suites counted once
← Gemini 3.8 FlashGemini 4 Hydrogen →
Data: Artificial Analysis, Vals AI, Surge AI and 20 othersatlas.village.dev

Margins

Each benchmark scaled 0–100% (its declared range, else the field’s range); Δ is Gemini 4 Hydrogen minus Gemini 3.8 Flash.

FrontierMath Tier 4 (v2) - PrivateMath22.087.8+65.9
TxBench-Antibody-Development V1 (60 evaluations)Science › Life sciences15.060.6+45.6
Code Migration v1.1 (Vals Index v2.2 subset)Software › Software engineering38.282.7+44.5
TaxCalcBench TY25Professional › Finance, accounting & tax6.044.0+38.0
Terminal-Bench 4.0 (Artificial Analysis)Software › Software engineering19.754.0+34.3
Terminal-Bench 4.0Software › Software engineering19.754.0+34.3
DAYJOB: Finance (October 2026 rescoring)Professional › Finance, accounting & tax4.538.5+34.0
FrontierSWE V2Software › Software engineering19.653.1+33.4
Terminal-Bench-Science 0.1Science › Multidisciplinary10.041.4+31.4
Humanity's Last Exam DiamondKnowledge & reasoning › General knowledge29.158.8+29.7
DAYJOB: AccountingProfessional › Finance, accounting & tax6.033.6+27.6
Furniture Assembly BenchmarkPerception & media › Spatial & visual reasoning31.757.5+25.8
ChartographyPerception & media › Documents & charts40.966.6+25.7
Ramp SWE-Bench v2 (Bash harness)Software › Software engineering47.070.8+23.8
LMArena Agent Arena Net Improvement (Steering Burden)Tool & computer use › General agents1.513.5+23.5
GDP.xlsxProfessional › General knowledge work18.340.9+22.6
DAYJOB: Healthcare (October 2026 rescoring)Health & medicine2.024.0+22.0
Humanity's Sixth SensePerception & media › Spatial & visual reasoning41.663.4+21.8
AutomationBench 1.0.6 (Zapier)Tool & computer use › Customer service & operations29.751.1+21.5
sudo L7 (October 2026 rescoring)Software › Software engineering10.631.7+21.1
LiveBench 2.0 (2026-11-06 release)Composite indexes57.377.0+19.7
FrontierMath Tiers 1-3 (v2) - PrivateMath68.488.1+19.6
KradleBenchGames24.343.9+19.6
TxBench Oligo Discovery V1 (113 evaluations)Science › Life sciences35.154.3+19.2
HANDBOOK.md Agents v1.1Tool & computer use › Customer service & operations26.544.6+18.1
Mystery Game PuzzlesGames47.065.0+18.0
FrontierCode 2.0 (Artificial Analysis)Software › Software engineering18.736.3+17.6
AA-Briefcase v1.1 EloProfessional › General knowledge work12041497+16.0
LMArena Code Arena WebDevSoftware › Frontend & design15851703+15.7
Legal Research BenchProfessional › Legal38.953.8+14.9
HiL-Bench (Human-in-Loop Benchmark)Tool & computer use › General agents41.555.3+13.9
AutomationBench-AA 1.0.6Tool & computer use › Customer service & operations60.974.4+13.5
CritPt v2 (agentic, avg@5)Science › Physical sciences65.178.0+12.9
LMArena Agent Arena Praise vs ComplaintTool & computer use › General agents4.029.3+12.6
Vals Index v2.2Composite indexes56.068.5+12.5
Tax Agent BenchProfessional › Finance, accounting & tax66.878.4+11.7
Medical Long Context Reasoning (MLCR-AA, October 2026 methodology)Health & medicine15.627.2+11.7
AA-Omniscience v1.1 IndexKnowledge & reasoning › General knowledge23.746.3+11.3
ComplexConstraints (November 2026 rescoring)Language & writing › Instruction following63.273.7+10.5
Artificial Analysis Intelligence Index v5.0.1Composite indexes30.841.2+10.4
GBENCH One-shot Coding GScoreGames56.666.8+10.2
LMArena Agent Arena Steering BurdenTool & computer use › General agents-2.817.4+10.1
GDP.pdf (November 2026 rescoring)Professional › General knowledge work44.254.2+10.0
GDP.pdf (Artificial Analysis, All-pass)Professional › General knowledge work22.832.8+10.0
Bug Hunt BenchSoftware › Software engineering17.127.0+9.8
DAYJOB: Law (November 2026 release)Professional › Legal0.09.7+9.7
SimpleBenchKnowledge & reasoning › General knowledge82.491.7+9.3
Vibe Code Bench v1.1Software › Frontend & design78.787.8+9.1
TxBench Antibody Discovery V1 (100 evaluations)Science › Life sciences49.057.7+8.7
LMArena Agent Arena Confirmed SuccessTool & computer use › General agents1.015.7+7.4
LMArena Agent Arena Bash ReliabilityTool & computer use › General agents8.9-4.5−6.7
TxBench Preclinical Pharmacology V1 Clean RerunsScience › Life sciences54.761.3+6.7
Epoch Capabilities Index (ECI)Composite indexes159167+6.4
micro1 PharmaBenchProfessional › General knowledge work5.611.4+5.8
Vals Finance IndexProfessional › Finance, accounting & tax60.566.1+5.7
Gray Swan IPI Attack Success (k=15)Safety & alignment5.50.0+5.5
FORTRESS Average Risk Score (ARS)Safety & alignment22.717.5+5.3
Professional Reasoning Benchmark - LegalProfessional › Legal48.953.1+4.2
Vals Legal IndexProfessional › Legal69.873.6+3.8
ARC-AGI-2Knowledge & reasoning › Abstract reasoning89.292.5+3.3
Excel Modeling Benchmark v1.1Professional › Finance, accounting & tax72.975.8+2.9
DAYJOB: TaxProfessional › Finance, accounting & tax0.83.6+2.8
Professional Reasoning Benchmark - FinanceProfessional › Finance, accounting & tax49.552.3+2.8
BuseyBench SVGPerception & media › Generation6.66.9+2.3
Chess Puzzles (task version 2.0)Games58.060.0+2.0
LMArena Text Arena EloLanguage & writing › General assistance14971490−1.2
SimpleQA Verified (Epoch anti-abstention)Knowledge & reasoning › General knowledge69.770.9+1.2
OTIS Mock AIME 2024-2025Math98.9100.0+1.1
HLAB-Vals-1.1Professional › Legal12.111.3−0.8
Finance Agent v2.1Professional › Finance, accounting & tax65.865.1−0.6
GradeBenchProfessional › Education51.551.9+0.4
LMArena Agent Arena Tool HallucinationTool & computer use › Tool calling0.00.7+0.3
ARC-AGI-1 Semi-PrivateKnowledge & reasoning › Abstract reasoning98.598.50.0
Dual-Arm Sim (task completion)Tool & computer use › Robotics0.00.00.0
DualEntry Accounting AI BenchmarkProfessional › Finance, accounting & tax83.283.20.0
Data: Artificial Analysis, Vals AI, Surge AI and 20 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…