Atlas

Compare Models

GPT-6.2 Astra vs GPT-6 Sol

GPT-6.2 Astra leads by 5.3 capability points (≈95% interval +2.6 to +8.1): a clear lead over GPT-6 Sol · P(higher) ≈ 100%

Wins 47 of 52 shared benchmarks (2 lost, 3 tied)

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

Each line runs through one model’s priced reasoning settings, cheapest to most expensive; grey points are other models’ priced settings in the same range. Results whose reasoning effort the source didn’t report are left out wherever a model has named effort levels. Capability and relative cost are the published estimates (relative cost is cost per task as a multiple of GPT-5 (high)); hover a setting, or use the arrow keys, to set it beside the other model’s same setting.

GPT-6.2 AstravsGPT-6 Solcapability vs. relative costData Nov 17, 2026Village LLM Atlasatlas.village.dev

Benchmark wins47 – 2 · 52 shared, 3 tied

Comparing matching setups only (2 benchmarks excluded) · suites counted once
← GPT-6.2 AstraGPT-6 Sol →

GPT-6.2 AstravsGPT-6 Solwins by categoryData Nov 17, 2026Data: Artificial Analysis, Epoch AI, Vals AI and 3 othersVillage LLM Atlasatlas.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 Astra.

ARC-AGI-3 (Semi-Private)General QA99.823.0−76.8
Terminal-Bench-Science 0.1Science78.130.0−48.1
Terminal-Bench Science 0.1 (Vals)Science71.430.0−41.4
EBR-bench v4 (card ban, single agent)Games92.453.3−39.0
Mystery Game PuzzlesGames95.056.0−39.0
SpireClimbGames41.39.2−32.2
AA-PersonalityBench Imagination indexChat & Writing60.429.6−31.2
Furniture Assembly BenchmarkMultimodal89.258.3−30.8
AA-PersonalityBench Conviction indexChat & Writing80.950.5−30.8
Humanity's Last Exam DiamondGeneral QA72.245.2−27.0
AA-PersonalityBench Expressiveness indexChat & Writing60.033.4−26.9
Vals Science IndexIndexes68.445.0−23.4
MysteryMechanismScience53.230.2−23.0
Chess Puzzles (task version 2.0)Games78.056.0−22.0
Terminal-Bench 4.0 (Artificial Analysis)Code64.143.9−20.2
AutomationBench 1.0.6 (Zapier)Agents52.433.2−19.2
MirrorCode (ML, +Private, 2L)Code67.848.9−18.9
Terminal-Bench 4.0Code63.144.4−18.7
Artificial Analysis Intelligence Index v5.0.1Indexes62.144.2−18.0
SimpleQA Verified (Epoch anti-abstention)General QA78.560.7−17.8
AutomationBench-AA 1.0.6Agents77.261.7−15.6
AA-Omniscience v1.1 IndexIndexes50.021.4−14.3
Legal Research BenchLegal42.828.8−13.9
AA-PersonalityBench Warmth indexChat & Writing67.854.3−13.6
AA-PersonalityBench Informativeness indexChat & Writing22.010.5−11.5
Harvey LAB-AA v1.1 (hallucination-gated)Legal15.03.6−11.4
FrontierMath Tier 4 (v2) - PrivateMath100.090.0−10.0
FrontierCode 2.0 (Artificial Analysis)Code45.435.6−9.8
AA-Briefcase v1.1 EloAgents16441476−9.2
FrontierMath ErdősMath10.31.5−8.8
Epoch Capabilities Index (ECI)Indexes174164−8.2
MMMU ProMultimodal89.582.5−6.9
HealthBench ProfessionalHealth & Medicine66.059.5−6.5
ARC-AGI-2General QA95.889.6−6.2
BioMysteryBench (Vals Terminus 2)Science80.774.8−5.9
Medical Long Context Reasoning (MLCR-AA, October 2026 methodology)Health & Medicine39.433.9−5.6
Tax Agent BenchFinance & Accounting57.953.0−4.8
GDP.pdf (Artificial Analysis, All-pass)Professional Work29.825.2−4.6
AA-PersonalityBench Organization indexChat & Writing16.612.4−4.3
AA-PersonalityBench Transparency indexChat & Writing36.139.1+3.0
ARC-AGI-1 Semi-PrivateGeneral QA98.595.5−3.0
CritPt v2 (agentic, avg@5)Science83.780.9−2.9
FrontierMath Tiers 1-3 (v2) - PrivateMath92.689.8−2.8
Vibe Code Bench v1.1Code89.987.8−2.1
MentalHealthBenchSafety53.354.2+0.9
Vals RSI Index v1.1 (four tasks)Agents29.028.1−0.9
CWE-Bench-AA (Artificial Analysis)Cybersecurity65.064.2−0.8
DeepsecBench-AA (Artificial Analysis)Cybersecurity46.846.3−0.5
Artificial Analysis Cyber Index v1Indexes37.336.8−0.4
AA-PersonalityBench Slop indexChat & Writing1.01.00.0
CyberGym-E2E-AA (Artificial Analysis)Cybersecurity0.00.00.0
OTIS Mock AIME 2024-2025Math100.0100.00.0

GPT-6.2 AstravsGPT-6 Solnormalised score marginsData Nov 17, 2026Data: Artificial Analysis, Epoch AI, Vals AI and 3 othersVillage LLM Atlasatlas.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…