Atlas

Compare Models

GPT-6.1 Terra vs GPT-6.2 Astra

GPT-6.2 Astra leads by 3.7 capability points (≈95% interval +0.9 to +6.5): a clear lead over GPT-6.1 Terra · P(higher) ≈ 100%

Wins 43 of 47 shared benchmarks (2 lost, 2 tied)

Capability vs. relative cost
GPT-6.1 TerraGPT-6.2 AstraClosed weightsOpen weights

Benchmark wins2 – 43 · 47 shared, 2 tied

Comparing matching setups only (3 benchmarks excluded) · suites counted once
← GPT-6.1 TerraGPT-6.2 Astra →
Data: Artificial Analysis, Epoch AI, Vals AI and 2 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-6.1 Terra.

KernelGen1P - OpenAI release percentageCode45.376.9+57.9
Humanity's Last Exam DiamondGeneral QA47.372.2+24.9
SimpleQA Verified (Epoch anti-abstention)General QA57.378.5+21.2
Mystery Game PuzzlesGames74.095.0+21.0
Terminal-Bench-Science 0.1Science57.678.1+20.5
SpireClimbGames22.341.3+19.0
AutomationBench-AA 1.0.6Agents60.277.2+17.1
Legal Research BenchLegal26.042.8+16.8
HLAB-Vals-1.1Legal1.716.3+14.6
EBR-bench v4 (card ban, single agent)Games78.192.4+14.3
PostTrainBench Lite - release-reported performanceCode60.574.8+14.3
Terminal-Bench 4.0Code49.563.1+13.6
AA-Briefcase v1.1 EloAgents14011643+13.3
Artificial Analysis Strategy & Ops IndexIndexes50.063.3+13.3
Artificial Analysis Economics IndexIndexes48.561.1+12.6
Artificial Analysis Legal IndexIndexes46.458.7+12.4
Artificial Analysis Healthcare & Medical IndexIndexes44.757.0+12.3
Artificial Analysis Finance & Accounting IndexIndexes49.060.9+11.9
Artificial Analysis Engineering IndexIndexes54.766.6+11.9
Code Migration v1.1 (Vals Index v2.2 subset)Code72.283.6+11.4
Vals Legal IndexIndexes58.869.7+10.9
Artificial Analysis Intelligence Index v5.0.1Indexes51.362.1+10.8
AA-Omniscience v1.1 IndexIndexes29.650.0+10.2
Furniture Assembly BenchmarkMultimodal79.289.2+10.0
MirrorCode (ML, +Private, 2L)Code76.767.8−8.9
Medical Long Context Reasoning (MLCR-AA, October 2026 methodology)Health & Medicine30.639.4+8.9
Terminal-Bench 4.0 (Artificial Analysis)Code55.664.1+8.6
FrontierMath Tier 4 (v2) - PrivateMath92.7100.0+7.3
Vals Index v2.2Indexes60.567.7+7.1
Vibe Code Bench v1.1Code84.089.9+5.9
OpenAI research debugging - internal release scoreCode76.481.8+5.5
Epoch Capabilities Index (ECI)Indexes168174+5.1
Artificial Analysis Coding Agent Index v2.1Indexes48.653.5+4.8
MMMU ProMultimodal84.789.5+4.8
Finance Agent v2.1Finance & Accounting56.060.7+4.8
Vals Finance IndexIndexes53.357.3+4.0
FrontierCode 2.0 (Artificial Analysis)Code42.345.4+3.1
Tax Agent BenchFinance & Accounting55.057.9+2.8
Excel Modeling Benchmark v1.1Finance & Accounting73.376.0+2.7
GPQA DiamondGeneral QA94.396.8+2.5
CritPt v2 (agentic, avg@5)Science81.483.7+2.3
MentalHealthBenchSafety55.553.3−2.2
HealthBench ProfessionalHealth & Medicine64.266.0+1.8
FrontierMath Tiers 1-3 (v2) - PrivateMath90.992.6+1.8
GDP.pdf (Artificial Analysis, All-pass)Professional Work28.629.8+1.2
DualEntry Accounting AI BenchmarkFinance & Accounting78.278.20.0
OTIS Mock AIME 2024-2025Math100.0100.00.0
Data: Artificial Analysis, Epoch AI, Vals AI and 2 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…