Atlas

Compare Models

GPT-5.6 Terra vs GPT-6.2 Astra

GPT-6.2 Astra leads by 8.1 capability points (≈95% interval +5.7 to +10.6): a clear lead over GPT-5.6 Terra · P(higher) ≈ 100%

Wins 52 of 54 shared benchmarks (2 lost)

Capability vs. relative cost
GPT-5.6 TerraGPT-6.2 AstraClosed weightsOpen weights

Benchmark wins2 – 52 · 54 shared

Comparing matching setups only (4 benchmarks excluded) · suites counted once
← GPT-5.6 TerraGPT-6.2 Astra →
Data: Artificial Analysis, Epoch AI, Arena and 11 othersatlas.village.dev

Margins

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

ARC-AGI-3 (Semi-Private)Games0.870.3+69.5
Terminal-Bench-Science 0.1Science › Multidisciplinary13.878.1+64.3
Terminal-Bench Science 0.1 (Vals)Science › Multidisciplinary10.071.4+61.4
Mystery Game PuzzlesGames35.095.0+60.0
ExploitBench - release-reported exploit progressSoftware › Cybersecurity46.2100.0+53.8
Humanity's Last Exam DiamondKnowledge & reasoning › General knowledge28.872.2+43.4
DAYJOB: Finance (October 2026 rescoring)Professional › Finance, accounting & tax3.345.8+42.5
DeepsecBench — Revalidated V1 (233 issues)Software › Cybersecurity19.161.2+42.2
MedChemBenchScience › Life sciences43.384.5+41.2
Arena Code WebDev — HTMLSoftware › Frontend & design15231827+40.7
Humanity's Sixth SensePerception & media › Spatial & visual reasoning22.763.2+40.5
Terminal-Bench 4.0Software › Software engineering22.763.1+40.4
AutomationBench 1.0.6 (Zapier)Tool & computer use › Customer service & operations14.852.4+37.6
Arena Code WebDev — ReactSoftware › Frontend & design15111784+37.5
Furniture Assembly BenchmarkPerception & media › Spatial & visual reasoning54.289.2+35.0
Gray Swan IPI Attack Success (k=15)Safety & alignment37.32.4+34.9
LMArena Agent Arena Net Improvement (Steering Burden)Tool & computer use › General agents-2.715.2+34.9
Chartography (November 2026 rescoring)Perception & media › Documents & charts58.890.7+31.9
Chess Puzzles (task version 2.0)Games47.078.0+31.0
DAYJOB: Healthcare (October 2026 rescoring)Health & medicine3.634.4+30.8
Blueprint-Bench 2Perception & media › Spatial & visual reasoning0.30.6+30.6
Artificial Analysis Intelligence Index v5.0.1Composite indexes32.362.1+29.8
AA-Omniscience v1.1 IndexKnowledge & reasoning › General knowledge-9.350.0+29.6
FrontierMath Tier 4 (v2) - PrivateMath70.7100.0+29.3
Bug Hunt BenchSoftware › Software engineering30.559.7+29.2
SEC-Bench Pro - release-reported cohortSoftware › Cybersecurity57.786.8+29.0
Terminal-Bench 4.0 (Artificial Analysis)Software › Software engineering35.464.1+28.8
MirrorCode (ML, +Private, 2L)Software › Software engineering41.167.8+26.7
DeepsecBench-AA (Artificial Analysis)Software › Cybersecurity25.146.8+21.7
FrontierCode 2.0 (Artificial Analysis)Software › Software engineering23.945.4+21.5
LMArena Agent Arena Praise vs ComplaintTool & computer use › General agents-3.337.0+20.2
LiveBench 2.0 (2026-11-06 release)Composite indexes67.284.9+17.7
AutomationBench-AA 1.0.6Tool & computer use › Customer service & operations59.677.2+17.6
AA-Briefcase v1.1 EloProfessional › General knowledge work13331644+17.0
Vibe Code Bench v1.1Software › Frontend & design74.689.9+15.3
CritPt v2 (agentic, avg@5)Science › Physical sciences69.783.7+14.0
ARC-AGI-2Knowledge & reasoning › Abstract reasoning82.595.8+13.3
Medical Long Context Reasoning (MLCR-AA, October 2026 methodology)Health & medicine27.239.4+12.2
LMArena Agent Arena Confirmed SuccessTool & computer use › General agents-7.712.4+10.1
MMMU ProKnowledge & reasoning › General knowledge79.589.5+10.0
Epoch Capabilities Index (ECI)Composite indexes161173+9.7
EnterpriseOps-Gym-AATool & computer use › Customer service & operations38.529.9−8.6
LMArena Agent Arena Steering BurdenTool & computer use › General agents-2.014.0+8.0
Tax Agent BenchProfessional › Finance, accounting & tax65.257.9−7.3
MentalHealthBenchHealth & medicine46.253.3+7.1
FrontierMath Tiers 1-3 (v2) - PrivateMath86.092.6+6.7
GDP.pdf (Artificial Analysis, All-pass)Professional › General knowledge work24.629.8+5.2
GPQA DiamondScience › Multidisciplinary92.997.0+4.1
LMArena Agent Arena Bash ReliabilityTool & computer use › General agents-0.37.7+4.0
HealthBench ProfessionalHealth & medicine62.466.0+3.6
ARC-AGI-1 Semi-PrivateKnowledge & reasoning › Abstract reasoning96.598.5+2.0
Legal Research BenchProfessional › Legal41.342.8+1.4
OTIS Mock AIME 2024-2025Math99.7100.0+0.3
LMArena Agent Arena Tool HallucinationTool & computer use › Tool calling1.21.20.0
Data: Artificial Analysis, Epoch AI, Arena and 11 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…