Atlas

Compare Models

GPT-6.2 Sol vs GPT-6 Sol

GPT-6.2 Sol leads by 3.4 capability points (≈95% interval +0.9 to +5.9): a clear lead over GPT-6 Sol · P(higher) ≈ 100%

Wins 53 of 62 shared benchmarks (8 lost, 1 tied)

Capability vs. relative cost
GPT-6.2 SolGPT-6 SolClosed weightsOpen weights

Benchmark wins53 – 8 · 62 shared, 1 tied

Comparing matching setups only (1 benchmark excluded) · suites counted once
← GPT-6.2 SolGPT-6 Sol →
Data: Artificial Analysis, Vals AI, Arena and 15 othersatlas.village.dev

Margins

Each benchmark scaled 0–100% (its declared range, else the field’s range); Δ is GPT-6 Sol minus GPT-6.2 Sol.

ARC-AGI-3 (Semi-Private)Games98.223.0−75.2
Terminal-Bench-Science 0.1Science › Multidisciplinary67.630.0−37.6
AA-PersonalityBench Organization indexLanguage & writing › Style & personality45.712.4−33.6
AA-PersonalityBench Imagination indexLanguage & writing › Style & personality60.129.6−30.8
AA-PersonalityBench Conviction indexLanguage & writing › Style & personality79.250.5−29.0
Terminal-Bench Science 0.1 (Vals)Science › Multidisciplinary58.630.0−28.6
AA-PersonalityBench Informativeness indexLanguage & writing › Style & personality37.710.5−27.5
Humanity's Sixth SensePerception & media › Spatial & visual reasoning56.831.2−25.6
Bug Hunt BenchSoftware › Software engineering52.127.9−24.1
AA-PersonalityBench Expressiveness indexLanguage & writing › Style & personality55.333.4−22.1
ChartographyPerception & media › Documents & charts75.153.6−21.5
Blueprint-Bench 2Perception & media › Spatial & visual reasoning0.60.4−20.1
Humanity's Last Exam DiamondKnowledge & reasoning › General knowledge64.045.2−18.8
AA-PersonalityBench Transparency indexLanguage & writing › Style & personality56.939.1−18.0
FrontierCUA v1 pass@1Tool & computer use › Computer & browser use25.07.5−17.5
AA-3D v1.0 EloSoftware › Frontend & design12591131−17.4
Terminal-Bench 4.0Software › Software engineering60.644.4−16.2
Terminal-Bench 4.0 (Artificial Analysis)Software › Software engineering60.143.9−16.2
AA-PersonalityBench Slop indexLanguage & writing › Style & personality15.41.0−14.5
DAYJOB: Finance (October 2026 rescoring)Professional › Finance, accounting & tax32.819.5−13.3
Artificial Analysis Intelligence Index v5.0.1Composite indexes57.344.2−13.1
Vals Science IndexScience › Multidisciplinary57.845.0−12.8
LMArena Code Arena WebDevSoftware › Frontend & design17571674−11.0
DAYJOB: Healthcare (October 2026 rescoring)Health & medicine17.27.2−10.0
Tax Agent BenchProfessional › Finance, accounting & tax62.953.0−9.9
AutomationBench-AA 1.0.6Tool & computer use › Customer service & operations71.361.7−9.6
AA-Omniscience v1.1 IndexKnowledge & reasoning › General knowledge40.421.4−9.5
LMArena Agent Arena Net Improvement (Steering Burden)Tool & computer use › General agents11.36.5−9.4
FrontierCode 2.0 (Artificial Analysis)Software › Software engineering44.435.6−8.8
Legal Research BenchProfessional › Legal37.028.8−8.2
AutomationBench 1.0.6 (Zapier)Tool & computer use › Customer service & operations41.333.2−8.0
DualEntry Accounting AI BenchmarkProfessional › Finance, accounting & tax84.276.2−8.0
Gray Swan IPI Attack Success (k=15)Safety & alignment2.310.1−7.8
Harvey LAB-AA v1.1 (hallucination-gated)Professional › Legal11.13.6−7.5
AA-Briefcase v1.1 EloProfessional › General knowledge work16121476−7.5
LMArena Agent Arena Praise vs ComplaintTool & computer use › General agents23.39.8−6.8
MysteryMechanismMath36.930.2−6.8
EnterpriseBench: CoreCraftTool & computer use › Customer service & operations36.443.1+6.7
DeepsecBench — Revalidated V1 (233 issues)Software › Cybersecurity47.340.7−6.6
HealthBench ProfessionalHealth & medicine65.559.5−6.0
AA-PersonalityBench Warmth indexLanguage & writing › Style & personality60.154.3−5.8
DeepsecBench-AA (Artificial Analysis)Software › Cybersecurity51.646.3−5.3
GDP.xlsxProfessional › General knowledge work22.317.1−5.2
GDP.pdf (Artificial Analysis, All-pass)Professional › General knowledge work30.025.2−4.8
Professional Reasoning Benchmark - LegalProfessional › Legal43.338.6−4.7
ARC-AGI-2Knowledge & reasoning › Abstract reasoning85.889.6+3.8
LMArena Agent Arena Confirmed SuccessTool & computer use › General agents10.93.9−3.5
Medical Long Context Reasoning (MLCR-AA, October 2026 methodology)Health & medicine37.233.9−3.3
MentalHealthBenchHealth & medicine50.954.2+3.3
SimpleBenchKnowledge & reasoning › General knowledge69.973.1+3.2
BioMysteryBench (Vals Terminus 2)Science › Life sciences77.874.8−3.0
LMArena Agent Arena Bash ReliabilityTool & computer use › General agents4.310.0+2.8
CWE-Bench-AA (Artificial Analysis)Software › Cybersecurity61.764.2+2.5
TxBench Preclinical Pharmacology V1 Clean RerunsScience › Life sciences65.363.0−2.3
CritPt v2 (agentic, avg@5)Science › Physical sciences78.680.9+2.3
LMArena Agent Arena Steering BurdenTool & computer use › General agents13.49.3−2.1
ARC-AGI-1 Semi-PrivateKnowledge & reasoning › Abstract reasoning97.595.5−2.0
TaxCalcBench TY25Professional › Finance, accounting & tax32.030.0−2.0
Artificial Analysis Cyber Index v1Software › Cybersecurity37.736.8−0.9
Vibe Code Bench v1.1Software › Frontend & design87.687.8+0.2
LMArena Agent Arena Tool HallucinationTool & computer use › Tool calling1.21.20.0
CyberGym-E2E-AA (Artificial Analysis)Software › Cybersecurity0.00.00.0
Data: Artificial Analysis, Vals AI, Arena and 15 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…