GPT-5.6 Sol vs GPT-6.2 Astra
GPT-6.2 Astra leads by 5.9 capability points (≈95% interval +3.4 to +8.5): a clear lead over GPT-5.6 Sol · P(higher) ≈ 100%
Wins 56 of 64 shared benchmarks (6 lost, 2 tied)
Capability vs. relative cost
GPT-5.6 SolGPT-6.2 AstraClosed weightsOpen weights
Benchmark wins6 – 56 · 64 shared, 2 tied
Comparing matching setups only (7 benchmarks excluded) · suites counted once
Area← GPT-5.6 SolGPT-6.2 Astra →
Data: Artificial Analysis, Epoch AI, Vals AI and 17 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 Sol.
| ARC-AGI-3 (Semi-Private) | Games | 7.8 | 70.3 | +62.5 |
| Terminal-Bench-Science 0.1 | Science › Multidisciplinary | 22.4 | 78.1 | +55.7 |
| Terminal-Bench Science 0.1 (Vals) | Science › Multidisciplinary | 20.0 | 71.4 | +51.4 |
| MirrorCode (ML, +Private, 2L) | Software › Software engineering | 20.0 | 67.8 | +47.8 |
| EBR-bench v4 (card ban, single agent) | Games | 44.8 | 92.4 | +47.6 |
| Humanity's Last Exam Diamond | Knowledge & reasoning › General knowledge | 30.4 | 72.2 | +41.8 |
| Mystery Game Puzzles | Games | 58.0 | 95.0 | +37.0 |
| DAYJOB: Finance (October 2026 rescoring) | Professional › Finance, accounting & tax | 11.8 | 45.8 | +34.0 |
| Humanity's Sixth Sense | Perception & media › Spatial & visual reasoning | 30.0 | 63.2 | +33.2 |
| Furniture Assembly Benchmark | Perception & media › Spatial & visual reasoning | 56.7 | 89.2 | +32.5 |
| DAYJOB: Healthcare (October 2026 rescoring) | Health & medicine | 4.4 | 34.4 | +30.0 |
| Blueprint-Bench 2 | Perception & media › Spatial & visual reasoning | 0.3 | 0.6 | +27.8 |
| Vals Science Index | Science › Multidisciplinary | 41.5 | 68.4 | +27.0 |
| DeepsecBench — Revalidated V1 (233 issues) | Software › Cybersecurity | 35.3 | 61.2 | +25.9 |
| Terminal-Bench 4.0 | Software › Software engineering | 37.9 | 63.1 | +25.3 |
| FrontierCUA v1 pass@1 | Tool & computer use › Computer & browser use | 7.5 | 32.5 | +25.0 |
| MedChemBench | Science › Life sciences | 59.8 | 84.5 | +24.7 |
| Gray Swan IPI Attack Success (k=15) | Safety & alignment | 27.0 | 2.4 | +24.6 |
| Terminal-Bench 4.0 (Artificial Analysis) | Software › Software engineering | 39.9 | 64.1 | +24.2 |
| AutomationBench 1.0.6 (Zapier) | Tool & computer use › Customer service & operations | 28.8 | 52.4 | +23.6 |
| LMArena Agent Arena Net Improvement (Steering Burden) | Tool & computer use › General agents | 3.7 | 15.2 | +22.5 |
| LMArena Code Arena WebDev | Software › Frontend & design | 1615 | 1781 | +22.0 |
| ExploitBench - release-reported exploit progress | Software › Cybersecurity | 78.5 | 100.0 | +21.5 |
| MysteryMechanism | Math | 33.3 | 53.2 | +19.8 |
| Chess Puzzles (task version 2.0) | Games | 59.0 | 78.0 | +19.0 |
| Bug Hunt Bench | Software › Software engineering | 41.4 | 59.7 | +18.3 |
| AutomationBench-AA 1.0.6 | Tool & computer use › Customer service & operations | 60.1 | 77.2 | +17.2 |
| FrontierMath Tier 4 (v2) - Private | Math | 82.9 | 100.0 | +17.1 |
| LMArena Agent Arena Praise vs Complaint | Tool & computer use › General agents | 3.8 | 37.0 | +16.6 |
| Agents' Last Exam v1.1 (ALE v1.1) | Professional › General knowledge work | 30.3 | 46.7 | +16.4 |
| AA-AnalystAgent (pass^5) | Professional › General knowledge work | 47.5 | 62.5 | +15.0 |
| EnterpriseOps-Gym-AA | Tool & computer use › Customer service & operations | 42.9 | 29.9 | −13.0 |
| Chartography (November 2026 rescoring) | Perception & media › Documents & charts | 78.6 | 90.7 | +12.1 |
| TaxCalcBench TY25 | Professional › Finance, accounting & tax | 32.0 | 44.0 | +12.0 |
| FrontierMath Erdős | Math | 0.0 | 10.3 | +10.3 |
| Tax Agent Bench | Professional › Finance, accounting & tax | 68.0 | 57.9 | −10.1 |
| BigFinanceBench private holdout (final-answer accuracy) | Professional › Finance, accounting & tax | 44.2 | 54.0 | +9.8 |
| BioMysteryBench (Vals Terminus 2) | Science › Life sciences | 71.1 | 80.7 | +9.6 |
| Medical Long Context Reasoning (MLCR-AA, October 2026 methodology) | Health & medicine | 30.0 | 39.4 | +9.4 |
| Vibe Code Bench v1.1 | Software › Frontend & design | 80.5 | 89.9 | +9.4 |
| AA-Briefcase v1.1 Elo | Professional › General knowledge work | 1479 | 1644 | +9.0 |
| DrugDiscoveryBench | Science › Life sciences | 56.5 | 65.5 | +9.0 |
| Epoch Capabilities Index (ECI) | Composite indexes | 163 | 173 | +7.9 |
| SEC-Bench Pro - release-reported cohort | Software › Cybersecurity | 79.1 | 86.8 | +7.7 |
| RiemannBench (November 2026 rescoring) | Math | 83.2 | 76.0 | −7.2 |
| LLM Debate Benchmark | Language & writing › Language & wordplay | 1625 | 1673 | +6.9 |
| MMMU Pro | Knowledge & reasoning › General knowledge | 82.7 | 89.5 | +6.8 |
| TxBench Preclinical Pharmacology V1 Clean Reruns | Science › Life sciences | 61.3 | 68.0 | +6.7 |
| MentalHealthBench | Health & medicine | 46.7 | 53.3 | +6.6 |
| CritPt v2 (agentic, avg@5) | Science › Physical sciences | 77.1 | 83.7 | +6.6 |
| Legal Research Bench | Professional › Legal | 48.1 | 42.8 | −5.3 |
| Vals RSI Index v1.1 (four tasks) | Software › Data & ML | 23.9 | 29.0 | +5.1 |
| LMArena Agent Arena Steering Burden | Tool & computer use › General agents | 3.8 | 14.0 | +5.1 |
| LMArena Agent Arena Confirmed Success | Tool & computer use › General agents | 2.5 | 12.4 | +5.0 |
| FrontierMath Tiers 1-3 (v2) - Private | Math | 89.1 | 92.6 | +3.5 |
| ARC-AGI-2 | Knowledge & reasoning › Abstract reasoning | 92.5 | 95.8 | +3.3 |
| EnterpriseBench: CoreCraft | Tool & computer use › Customer service & operations | 46.2 | 43.6 | −2.6 |
| GPQA Diamond | Science › Multidisciplinary | 94.6 | 97.0 | +2.4 |
| GDP.pdf (Artificial Analysis, All-pass) | Professional › General knowledge work | 27.8 | 29.8 | +2.0 |
| HealthBench Professional | Health & medicine | 64.1 | 66.0 | +1.9 |
| LMArena Agent Arena Bash Reliability | Tool & computer use › General agents | 10.5 | 7.7 | −1.4 |
| ARC-AGI-1 Semi-Private | Knowledge & reasoning › Abstract reasoning | 97.5 | 98.5 | +1.0 |
| LMArena Agent Arena Tool Hallucination | Tool & computer use › Tool calling | 1.2 | 1.2 | 0.0 |
| OTIS Mock AIME 2024-2025 | Math | 100.0 | 100.0 | 0.0 |
Data: Artificial Analysis, Epoch AI, Vals AI and 17 othersatlas.village.dev