Atlas

Compare Models

Gemini 4 Neon vs GPT-6 Sol

Gemini 4 Neon leads by 5.4 capability points (≈95% interval +2.6 to +8.1): a clear lead over GPT-6 Sol · P(higher) ≈ 100%

Wins 41 of 46 shared benchmarks (4 lost, 1 tied)

Capability vs. relative cost
Gemini 4 NeonGPT-6 SolClosed weightsOpen weights

Benchmark wins41 – 4 · 46 shared, 1 tied

Comparing matching setups only (13 benchmarks excluded) · suites counted once
← Gemini 4 NeonGPT-6 Sol →
Data: Artificial Analysis, Vals AI, Epoch AI and 9 othersatlas.village.dev

Margins

Each benchmark scaled 0–100% (its declared range, else the field’s range); Δ is GPT-6 Sol minus Gemini 4 Neon.

Terminal-Bench-Science 0.1Science › Multidisciplinary70.017.6−52.4
Humanity's Sixth SensePerception & media › Spatial & visual reasoning75.931.2−44.7
DAYJOB: Healthcare (October 2026 rescoring)Health & medicine47.27.2−40.0
EnterpriseBench: CoreCraftTool & computer use › Customer service & operations77.443.1−34.3
Terminal-Bench 4.0 (Artificial Analysis)Software › Software engineering63.630.3−33.3
Terminal-Bench Science 0.1 (Vals)Science › Multidisciplinary62.930.0−32.9
DAYJOB: Finance (October 2026 rescoring)Professional › Finance, accounting & tax50.519.5−31.0
Furniture Assembly BenchmarkPerception & media › Spatial & visual reasoning89.258.3−30.8
Mystery Game PuzzlesGames85.056.0−29.0
Legal Research BenchProfessional › Legal56.328.8−27.4
Chess Puzzles (task version 2.0)Games83.056.0−27.0
Tax Agent BenchProfessional › Finance, accounting & tax79.853.0−26.7
Humanity's Last Exam DiamondKnowledge & reasoning › General knowledge71.345.2−26.1
GDP.xlsxProfessional › General knowledge work43.117.1−26.0
LMArena Agent Arena Net Improvement (Steering Burden)Tool & computer use › General agents18.46.5−23.3
AutomationBench 1.0.6 (Zapier)Tool & computer use › Customer service & operations56.333.2−23.1
Harvey LAB (Mercor run, undisclosed ~213-task cohort)Professional › Legal26.64.1−22.5
SimpleBenchKnowledge & reasoning › General knowledge94.473.1−21.3
APEX-Agents 1.1Professional › General knowledge work88.968.7−20.2
ChartographyPerception & media › Documents & charts73.553.6−19.9
AA-Briefcase v1.1 EloProfessional › General knowledge work16991359−18.7
AutomationBench-AA 1.0.6Tool & computer use › Customer service & operations79.361.7−17.6
Vals Science IndexScience › Multidisciplinary62.645.0−17.6
LMArena Agent Arena Praise vs ComplaintTool & computer use › General agents44.59.8−17.4
Artificial Analysis Intelligence Index v5.0.1Composite indexes61.444.2−17.2
AA-Omniscience v1.1 IndexKnowledge & reasoning › General knowledge54.821.4−16.7
τ³-Banking (Mercor run, bm25_grep retrieval)Tool & computer use › Customer service & operations59.543.0−16.5
MysteryMechanismMath45.030.2−14.9
SimpleQA Verified (Epoch anti-abstention)Knowledge & reasoning › General knowledge75.260.7−14.5
Vending-Bench 2Business simulations1684314428−14.3
Terminal-Bench 4.0Software › Software engineering61.649.4−12.2
APEX-Accounting 1.1Professional › Finance, accounting & tax70.859.7−11.1
LMArena Text Arena EloLanguage & writing › General assistance15181458−10.6
FrontierCode 2.0 (Artificial Analysis)Software › Software engineering44.435.6−8.8
GDP.pdf (Artificial Analysis, All-pass)Professional › General knowledge work33.425.2−8.2
LMArena Agent Arena Bash ReliabilityTool & computer use › General agents-6.010.0+8.0
LMArena Agent Arena Confirmed SuccessTool & computer use › General agents19.23.9−7.7
FrontierMath Tier 4 (v2) - PrivateMath97.690.0−7.6
LMArena Agent Arena Steering BurdenTool & computer use › General agents21.29.3−6.0
Vibe Code Bench v1.1Software › Frontend & design93.787.8−5.9
BioMysteryBench (Vals Terminus 2)Science › Life sciences80.074.8−5.2
CritPt v2 (agentic, avg@5)Science › Physical sciences78.080.9+2.9
Medical Long Context Reasoning (MLCR-AA, October 2026 methodology)Health & medicine31.733.9+2.2
FrontierMath Tiers 1-3 (v2) - PrivateMath91.289.8−1.4
LMArena Agent Arena Tool HallucinationTool & computer use › Tool calling1.01.2+0.1
OTIS Mock AIME 2024-2025Math100.0100.00.0
Data: Artificial Analysis, Vals AI, Epoch AI and 9 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…