GPT-5.4 vs Muse Spark 1.1
GPT-5.4 
Muse Spark 1.1 Verdict
- Muse Spark 1.1 leads on quality: 64.9 vs 63.4.
- GPT-5.4 is stronger in agents & tools, composite, long context.
- Muse Spark 1.1 is stronger in coding, human preference, instruction following, knowledge, multimodal, reasoning, maths.
- Muse Spark 1.1 is 2.8× cheaper ($2.00 vs $5.63 per 1M blended).
- Muse Spark 1.1 streams 1.5× faster (194 vs 130 tokens per second).
| Metric | GPT-5.4 | Muse Spark 1.1 |
|---|---|---|
| BenchLeader Index | 63.4 | 64.9 |
| Agents & tools score | 66.1 | 62.7 |
| Coding score | 51.6 | 68.2 |
| Composite score | 78.1 | 72.2 |
| Human preference score | 64.5 | 66.0 |
| Instruction following score | 71.7 | 77.8 |
| Knowledge score | 66.8 | 73.1 |
| Long context score | 67.6 | 65.3 |
| Multimodal score | 64.0 | 64.9 |
| Reasoning score | 62.3 | 68.8 |
| Maths score | – | 52.2 |
| Blended price $/M | $5.63 | $2.00 |
| Output speed | 130 tok/s | 194 tok/s |
| Time to first answer | 126.4 s | 2.0 s |
| Context window | 1.1M | 1.0M |
| SimpleQA Verified | – | 57.8% |
| Terminal-Bench | 81.8% | – |
| SciCode | – | 58.2% |
| APEX-Agents | 34.9% | 41.9% |
| ProofBench | – | 39.0% |
| Epoch Capabilities Index | 156.9 | 154.6 |
| LMArena Text | 1466 | 1492 |
| LMArena Hard Prompts | 1488 | 1511 |
| LMArena Coding | 1514 | 1531 |
| LMArena WebDev | 1386 | 1541 |
| LMArena Vision | 1293 | 1293 |
| LMArena Agent | – | -2.6 |
| AA Intelligence Index | 39.0 | 34.3 |
| IFBench | 74.0% | – |
| AA-LCR | 82.0% | 77.7% |
| MMMU-Pro | 78.4% | – |
| AA-Omniscience | 5.8 | 28.1 |
| Terminal-Bench Hard | 57.6% | – |
| GPQA Diamond (AA) | 92.0% | 89.8% |
| Humanity's Last Exam (AA) | 43.7% | 46.2% |
| SciCode (AA) | – | 58.8% |
| τ²-Bench Telecom (AA) | 87.1% | – |
| SWE-Bench Pro | – | 61.5% |
| MCP Atlas | – | 88.1% |
| MultiChallenge | – | 75.3% |
| PRBench Finance | – | 55.0% |
| PRBench Legal | – | 57.0% |
| MultiNRC | – | 65.6% |
| HiL-Bench | 9.7% | – |
| EQ-Bench 4 | 1272 | 1260 |
| Kagi LLM Benchmark | 63.8% | – |
Data as of 2026-09-09. Best configuration of each model; every score links to its source on the model pages.