Compare / DeepSeek V4.1 Flash vs GPT-6 Luna
DeepSeek V4.1 Flash vs GPT-6 Luna
46 results from 3 sources that measured both models. Each row is in the source's own units; there is no overall winner.
| Benchmark | Metric | DeepSeek V4.1 Flash | GPT-6 Luna |
|---|---|---|---|
| CatalogBenchmeasured by us | Channel rules broken% of products | 1.8% | 0.0% |
| CatalogBenchmeasured by us | Claims to check% of products | 0.0% | 0.0% |
| CatalogBenchmeasured by us | Content quality% of checks | 96.2% | 97.4% |
| CatalogBenchmeasured by us | Failed outputs% of products | 0.6% | 0.0% |
| CatalogBenchmeasured by us | Missing UK information% of products | 0.0% | 0.0% |
| CatalogBenchmeasured by us | Not findable% of products | 3.6% | 10.1% |
| CatalogBenchmeasured by us | Publish-ready listings% of products | 47.0% | 67.9% |
| CatalogBenchmeasured by us | Reliably publish-ready% of products | 23.2% | 51.8% |
| CatalogBenchmeasured by us | Unsupported claimsclaims per product | 1.25 | 0.34 |
| CatalogBenchmeasured by us | Unsupported claims% of products | 39.9% | 8.3% |
| CatalogBenchmeasured by us | Wrong attributes% of products | 20.2% | 19.1% |
| CatalogBenchmeasured by us | Wrong category or variant% of products | 2.4% | 0.0% |
| CatalogBenchmeasured by us | Channel compliance% of products | 99.4% | 100.0% |
| CatalogBenchmeasured by us | Channel rules broken% of products | 3.0% | 0.0% |
| CatalogBenchmeasured by us | Claims to check% of products | 0.0% | 0.0% |
| CatalogBenchmeasured by us | Conflicts caught% of conflicts | 100.0% | 100.0% |
| CatalogBenchmeasured by us | Content quality% of checks | 96.7% | 97.0% |
| CatalogBenchmeasured by us | Cost per productUS dollars | $0.0061 | $0.0005 |
| CatalogBenchmeasured by us | Decision accuracy% of decisions | 96.9% | 97.0% |
| CatalogBenchmeasured by us | Failed outputs% of products | 0.0% | 0.0% |
| CatalogBenchmeasured by us | Field accuracy% of missing fields | 94.4% | 93.7% |
| CatalogBenchmeasured by us | Invented values% of filled values | 3.1% | 1.7% |
| CatalogBenchmeasured by us | Missing UK information% of products | 0.0% | 0.0% |
| CatalogBenchmeasured by us | Not findable% of products | 3.6% | 10.1% |
| CatalogBenchmeasured by us | Publish-ready listings% of products | 63.7% | 67.9% |
| CatalogBenchmeasured by us | Reliably publish-ready% of products | 41.1% | 55.4% |
| CatalogBenchmeasured by us | Unsupported claimsclaims per product | 0.54 | 0.31 |
| CatalogBenchmeasured by us | Unsupported claims% of products | 13.1% | 8.9% |
| CatalogBenchmeasured by us | Wrong attributes% of products | 20.2% | 18.4% |
| CatalogBenchmeasured by us | Wrong category or variant% of products | 3.0% | 1.8% |
| Arena (formerly LMArena)reported | Confirmed task successIPS effect estimate | 0.12 | 0.05 |
| Arena (formerly LMArena)reported | Praise over complaintIPS effect estimate | 0.03 | -0.03 |
| Arena (formerly LMArena)reported | SteerabilityIPS effect estimate | -0.03 | 0.02 |
| Arena (formerly LMArena)reported | Tool groundingIPS effect estimate | 0.00 | 0.00 |
| Arena (formerly LMArena)reported | OverallArena rating | 1,471 | 1,422 |
| Arena (formerly LMArena)reported | Business, management and financeArena rating | 1,481 | 1,452 |
| Arena (formerly LMArena)reported | Creative writingArena rating | 1,440 | 1,403 |
| Arena (formerly LMArena)reported | Expert promptsArena rating | 1,518 | 1,505 |
| Arena (formerly LMArena)reported | Instruction followingArena rating | 1,478 | 1,443 |
| Arena (formerly LMArena)reported | OverallArena rating | 1,477 | 1,442 |
| Arena (formerly LMArena)reported | Writing, literature and languageArena rating | 1,455 | 1,415 |
| Artificial Analysisreported | Artificial Analysis Intelligence Indexindex score | 39.5 | 38.1 |
| Artificial Analysisreported | Humanity's Last Exam% of questions | 39.2% | 38.5% |
| Artificial Analysisreported | Long-context reasoning (AA-LCR)% of questions | 84.0% | 83.3% |
| Artificial Analysisreported | SciCode% of problems | 51.9% | 54.6% |
| Artificial Analysisreported | Terminal-Bench 4.0% of tasks | 26.8% | 12.6% |
Bold green marks the better value on that metric. Where a source reports ranges that overlap, the difference may not be meaningful; see the benchmark page for ranges.