GPT-5.5 vs GPT-5.6 Sol
GPT-5.5 Verdict
- GPT-5.6 Sol (high) leads on quality: 68.0 vs 66.6.
- GPT-5.5 is stronger in human preference, instruction following, long context, reasoning.
- GPT-5.6 Sol (high) is stronger in agents & tools, coding, composite, multimodal.
- GPT-5.6 Sol (high) is 1.4× cheaper ($8.00 vs $11.25 per 1M blended).
- GPT-5.5 streams 1.5× faster (88 vs 59 tokens per second).
| Metric | GPT-5.5 | GPT-5.6 Sol (high) |
|---|---|---|
| BenchLeader Index | 66.6 | 68.0 |
| Agents & tools score | 68.4 | 87.6 |
| Coding score | 57.3 | 72.0 |
| Composite score | 77.7 | 82.5 |
| Human preference score | 67.6 | – |
| Instruction following score | 73.3 | 67.6 |
| Knowledge score | 73.7 | 73.7 |
| Long context score | 68.8 | 67.4 |
| Multimodal score | 64.9 | 66.5 |
| Reasoning score | 69.7 | 62.9 |
| Blended price $/M | $11.25 | $8.00 |
| Output speed | 88 tok/s | 59 tok/s |
| Time to first answer | 62.3 s | 31.4 s |
| Context window | 1.1M | 1M |
| Terminal-Bench | 84.7% | – |
| SimpleBench | 69.0% | – |
| SciCode | – | 56.9% |
| Remote Labor Index | 6.3% | – |
| WeirdML | – | 88.8% |
| APEX-Agents | 38.5% | – |
| FrontierCode | 43.0% | – |
| Epoch Capabilities Index | 159.1 | – |
| LMArena Text | 1477 | – |
| LMArena Hard Prompts | 1498 | – |
| LMArena Coding | 1509 | – |
| LMArena WebDev | 1458 | – |
| LMArena Vision | 1296 | – |
| LMArena Agent | 3 | – |
| AA Intelligence Index | 38.6 | 42.5 |
| IFBench | 75.8% | 69.2% |
| AA-LCR | 84.3% | 81.7% |
| MMMU-Pro | 79.9% | 81.8% |
| AA-Omniscience | 20.5 | 20.4 |
| Terminal-Bench Hard | 60.6% | 62.1% |
| GPQA Diamond (AA) | 93.5% | 92.8% |
| Humanity's Last Exam (AA) | 45.8% | 46.0% |
| SciCode (AA) | 55.8% | 57.8% |
| τ²-Bench Telecom (AA) | 93.9% | 83.3% |
| EnigmaEval | – | 37.1% |
| HiL-Bench | 39.7% | – |
| EQ-Bench 4 | 1315 | – |
| Kagi LLM Benchmark | 88.8% | – |
| ARC-AGI-1 | – | 97.0% |
| ARC-AGI-2 | – | 85.4% |
| ARC-AGI-3 | – | 2.2% |
Data as of 2026-09-09. Best configuration of each model; every score links to its source on the model pages.