Atlas

Compare Models

GPT-5.6 Sol vs GPT-6.2 Sol

GPT-6.2 Sol leads by 4.3 capability points (≈95% interval +1.9 to +6.8): a clear lead over GPT-5.6 Sol · P(higher) ≈ 100%

Wins 38 of 48 shared benchmarks (7 lost, 3 tied)

Capability vs. relative cost
GPT-5.6 SolGPT-6.2 SolClosed weightsOpen weights

Benchmark wins7 – 38 · 48 shared, 3 tied

Comparing matching setups only (9 benchmarks excluded) · suites counted once
← GPT-5.6 SolGPT-6.2 Sol →
Data: Artificial Analysis, Vals AI, Arena and 16 othersatlas.village.dev

Margins

Each benchmark scaled 0–100% (its declared range, else the field’s range); Δ is GPT-6.2 Sol minus GPT-5.6 Sol.

ARC-AGI-3 (Semi-Private)Games7.854.9+47.1
Terminal-Bench-Science 0.1Science › Multidisciplinary22.467.6+45.2
Terminal-Bench Science 0.1 (Vals)Science › Multidisciplinary20.058.6+38.6
Humanity's Last Exam DiamondKnowledge & reasoning › General knowledge30.464.0+33.6
ChartographyPerception & media › Documents & charts45.075.1+30.1
Humanity's Sixth SensePerception & media › Spatial & visual reasoning30.056.8+26.8
Gray Swan IPI Attack Success (k=15)Safety & alignment27.02.3+24.7
Blueprint-Bench 2Perception & media › Spatial & visual reasoning0.30.6+23.4
Terminal-Bench 4.0Software › Software engineering37.960.6+22.7
DAYJOB: Finance (October 2026 rescoring)Professional › Finance, accounting & tax11.832.8+21.0
Terminal-Bench 4.0 (Artificial Analysis)Software › Software engineering39.960.1+20.2
LMArena Code Arena WebDevSoftware › Frontend & design16151757+18.8
FrontierCUA v1 pass@1Tool & computer use › Computer & browser use7.525.0+17.5
Vals Science IndexScience › Multidisciplinary41.557.8+16.3
LMArena Agent Arena Net Improvement (Steering Burden)Tool & computer use › General agents3.711.3+14.9
EnterpriseOps-Gym-AATool & computer use › Customer service & operations42.928.9−14.0
DAYJOB: Healthcare (October 2026 rescoring)Health & medicine4.417.2+12.8
AutomationBench 1.0.6 (Zapier)Tool & computer use › Customer service & operations28.841.3+12.5
DeepsecBench — Revalidated V1 (233 issues)Software › Cybersecurity35.347.3+12.0
AutomationBench-AA 1.0.6Tool & computer use › Customer service & operations60.171.3+11.2
Legal Research BenchProfessional › Legal48.137.0−11.1
Bug Hunt BenchSoftware › Software engineering41.452.1+10.6
Agents' Last Exam v1.1 (ALE v1.1)Professional › General knowledge work30.340.2+9.8
EnterpriseBench: CoreCraftTool & computer use › Customer service & operations46.236.4−9.8
LMArena Agent Arena Praise vs ComplaintTool & computer use › General agents3.823.3+9.7
AA-Briefcase v1.1 EloProfessional › General knowledge work14791612+7.3
Medical Long Context Reasoning (MLCR-AA, October 2026 methodology)Health & medicine30.037.2+7.2
Vibe Code Bench v1.1Software › Frontend & design80.587.6+7.1
SWE-sweepSoftware › Software engineering4.711.6+6.9
BioMysteryBench (Vals Terminus 2)Science › Life sciences71.177.8+6.7
ARC-AGI-2Knowledge & reasoning › Abstract reasoning92.585.8−6.7
AA-AnalystAgent (pass^5)Professional › General knowledge work47.553.8+6.3
DrugDiscoveryBenchScience › Life sciences56.562.2+5.7
Tax Agent BenchProfessional › Finance, accounting & tax68.062.9−5.1
LMArena Agent Arena Steering BurdenTool & computer use › General agents3.813.4+4.8
MentalHealthBenchHealth & medicine46.750.9+4.2
LMArena Agent Arena Confirmed SuccessTool & computer use › General agents2.510.9+4.2
TxBench Preclinical Pharmacology V1 Clean RerunsScience › Life sciences61.365.3+4.0
MysteryMechanismMath33.336.9+3.6
RiemannBench (November 2026 rescoring)Math83.280.0−3.2
LMArena Agent Arena Bash ReliabilityTool & computer use › General agents10.54.3−3.1
GDP.pdf (Artificial Analysis, All-pass)Professional › General knowledge work27.830.0+2.2
CritPt v2 (agentic, avg@5)Science › Physical sciences77.178.6+1.4
GPQA DiamondScience › Multidisciplinary94.696.0+1.4
HealthBench ProfessionalHealth & medicine64.165.5+1.4
LMArena Agent Arena Tool HallucinationTool & computer use › Tool calling1.21.20.0
ARC-AGI-1 Semi-PrivateKnowledge & reasoning › Abstract reasoning97.597.50.0
TaxCalcBench TY25Professional › Finance, accounting & tax32.032.00.0
Data: Artificial Analysis, Vals AI, Arena and 16 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…