Atlas

Compare Models

Claude Opus 5 vs GPT-6.2 Sol

GPT-6.2 Sol leads by 3.0 capability points (≈95% interval +0.5 to +5.6): a clear lead over Claude Opus 5 · P(higher) ≈ 99%

Wins 29 of 43 shared benchmarks (12 lost, 2 tied)

Capability vs. relative cost
Claude Opus 5GPT-6.2 SolClosed weightsOpen weights

Benchmark wins12 – 29 · 43 shared, 2 tied

Comparing matching setups only (11 benchmarks excluded) · suites counted once
← Claude Opus 5GPT-6.2 Sol →
Data: Artificial Analysis, Vals AI, Arena and 12 othersatlas.village.dev

Margins

Each benchmark scaled 0–100% (its declared range, else the field’s range); Δ is GPT-6.2 Sol minus Claude Opus 5.

Terminal-Bench-Science 0.1Science › Multidisciplinary28.667.6+39.0
EnterpriseBench: CoreCraftTool & computer use › Customer service & operations68.736.4−32.3
Terminal-Bench Science 0.1 (Vals)Science › Multidisciplinary27.158.6+31.4
ExploitBench - release-reported exploit progressSoftware › Cybersecurity70.0100.0+30.0
Blueprint-Bench 2Perception & media › Spatial & visual reasoning0.30.6+26.6
Humanity's Sixth SensePerception & media › Spatial & visual reasoning30.556.8+26.3
DeepsecBench-AA (Artificial Analysis)Software › Cybersecurity25.951.6+25.7
Medical Long Context Reasoning (MLCR-AA, October 2026 methodology)Health & medicine13.937.2+23.3
Humanity's Last Exam DiamondKnowledge & reasoning › General knowledge44.864.0+19.2
EnterpriseOps-Gym-AATool & computer use › Customer service & operations47.528.9−18.6
Legal Research BenchProfessional › Legal55.337.0−18.3
DeepsecBench — Revalidated V1 (233 issues)Software › Cybersecurity32.347.3+15.0
AutomationBench-AA 1.0.6Tool & computer use › Customer service & operations56.671.3+14.7
AutomationBench 1.0.6 (Zapier)Tool & computer use › Customer service & operations26.941.3+14.3
Tax Agent BenchProfessional › Finance, accounting & tax76.762.9−13.7
FrontierCode 2.0 (Artificial Analysis)Software › Software engineering31.244.4+13.2
Artificial Analysis Intelligence Index v5.0.1Composite indexes44.257.3+13.1
TaxCalcBench TY25Professional › Finance, accounting & tax20.032.0+12.0
Terminal-Bench 4.0 (Artificial Analysis)Software › Software engineering49.060.1+11.1
SimpleBenchKnowledge & reasoning › General knowledge80.669.9−10.7
SWE-sweepSoftware › Software engineering1.311.6+10.3
LMArena Agent Arena Net Improvement (Steering Burden)Tool & computer use › General agents6.311.3+9.8
Vals Science IndexScience › Multidisciplinary47.957.8+9.8
DAYJOB: Finance (October 2026 rescoring)Professional › Finance, accounting & tax23.832.8+9.0
GDP.pdf (Artificial Analysis, All-pass)Professional › General knowledge work21.630.0+8.4
CritPt v2 (agentic, avg@5)Science › Physical sciences71.478.6+7.1
LMArena Agent Arena Praise vs ComplaintTool & computer use › General agents10.323.3+6.5
LMArena Agent Arena Steering BurdenTool & computer use › General agents2.413.4+5.5
DAYJOB: Healthcare (October 2026 rescoring)Health & medicine22.017.2−4.8
ARC-AGI-2Knowledge & reasoning › Abstract reasoning90.485.8−4.6
Chartography (November 2026 rescoring)Perception & media › Documents & charts84.487.7+3.3
TxBench Preclinical Pharmacology V1 Clean RerunsScience › Life sciences63.065.3+2.3
AA-Briefcase v1.1 EloProfessional › General knowledge work16541612−2.3
LMArena Agent Arena Bash ReliabilityTool & computer use › General agents8.74.3−2.2
BioMysteryBench (Vals Terminus 2)Science › Life sciences79.377.8−1.5
AA-Omniscience v1.1 IndexKnowledge & reasoning › General knowledge37.640.4+1.4
ARC-AGI-3 (Semi-Private)Games30.231.1+0.9
Vibe Code Bench v1.1Software › Frontend & design88.487.6−0.8
MysteryMechanismMath37.436.9−0.5
LMArena Agent Arena Confirmed SuccessTool & computer use › General agents10.110.9+0.4
LMArena Agent Arena Tool HallucinationTool & computer use › Tool calling1.01.2+0.1
AA-AnalystAgent (pass^5)Professional › General knowledge work53.853.80.0
ARC-AGI-1 Semi-PrivateKnowledge & reasoning › Abstract reasoning97.597.50.0
Data: Artificial Analysis, Vals AI, Arena and 12 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…