GPT-5.6 Sol vs GPT-6 Astra
GPT-5.6 Sol Verdict
- GPT-6 Astra (max) leads on quality: 70.6 vs 68.0.
- GPT-5.6 Sol (high) is stronger in agents & tools, composite, instruction following, long context, multimodal.
- GPT-6 Astra (max) is stronger in coding, knowledge, reasoning, maths.
- GPT-5.6 Sol (high) is 2.8× cheaper ($8.00 vs $22.00 per 1M blended).
| Metric | GPT-5.6 Sol (high) | GPT-6 Astra (max) |
|---|---|---|
| BenchLeader Index | 68.0 | 70.6 |
| Agents & tools score | 87.6 | 68.2 |
| Coding score | 72.0 | 79.0 |
| Composite score | 82.5 | 73.1 |
| Instruction following score | 67.6 | – |
| Knowledge score | 73.7 | 79.8 |
| Long context score | 67.4 | – |
| Multimodal score | 66.5 | – |
| Reasoning score | 62.9 | 70.9 |
| Maths score | – | 77.0 |
| Blended price $/M | $8.00 | $22.00 |
| Output speed | 59 tok/s | 54 tok/s |
| Time to first answer | 31.4 s | 328.6 s |
| Context window | 1M | 1.1M |
| GPQA Diamond | – | 95.8% |
| FrontierMath Tiers 1–3 | – | 93.7% |
| FrontierMath Tier 4 | – | 97.6% |
| OTIS Mock AIME | – | 100.0% |
| SimpleQA Verified | – | 75.6% |
| Terminal-Bench | – | 58.2% |
| SciCode | 56.9% | 56.5% |
| WeirdML | 88.8% | – |
| FrontierCode | – | 53.3% |
| LMArena WebDev | – | 1796 |
| LMArena Agent | – | 12.5 |
| LiveBench | – | 82.2% |
| LiveBench Reasoning | – | 92.7% |
| LiveBench Coding | – | 80.4% |
| LiveBench Agentic Coding | – | 57.3% |
| LiveBench Mathematics | – | 96.8% |
| LiveBench Data Analysis | – | 83.0% |
| LiveBench Language | – | 89.4% |
| AA Intelligence Index | 42.5 | – |
| IFBench | 69.2% | – |
| AA-LCR | 81.7% | – |
| MMMU-Pro | 81.8% | – |
| AA-Omniscience | 20.4 | – |
| Terminal-Bench Hard | 62.1% | – |
| GPQA Diamond (AA) | 92.8% | – |
| Humanity's Last Exam (AA) | 46.0% | – |
| SciCode (AA) | 57.8% | – |
| τ²-Bench Telecom (AA) | 83.3% | – |
| Terminal-Bench 2.1 (Vals) | – | 87.3% |
| Vals Index | – | 66.6 |
| EnigmaEval | 37.1% | – |
| ARC-AGI-1 | 97.0% | 97.5% |
| ARC-AGI-2 | 85.4% | 95.0% |
| ARC-AGI-3 | 2.2% | 62.7% |
Data as of 2026-09-09. Best configuration of each model; every score links to its source on the model pages.