Up to four models
DeepSeek-V4.1-Flash vs GPT-5.6 Luna
On the index (GPT-4 = 100), DeepSeek-V4.1-Flash scores 167 and GPT-5.6 Luna 167. Side by side on the same third-party tests, with price and context. The link in your address bar shares this exact comparison.
DeepSeek-V4.1-Flash167151–184
WeightsOpen
ReleasedSep 2026
Price in / out$0.14 / $0.42
Context1.05M
Core tests taken6 of 9
GPT-5.6 Luna
167151–184
WeightsClosed
ReleasedJul 2026
Price in / out$0.20 / $1.2
Context1.05M
Core tests taken6 of 9
DeepSeek-V4.1-Flash and GPT-5.6 Luna are too close to call (0 points apart, within the uncertainty), ahead on 2 of 6 shared core tests, and costs 2.1× less.
Test by testthird-party results
DeepSeek-V4.1-FlashGPT-5.6 LunaEvery result as a table
| Test | DeepSeek-V4.1-Flash | GPT-5.6 Luna |
|---|---|---|
| Reasoning & knowledge | ||
| GPQA Diamond (Epoch AI) | 89.8% | 91.6% |
| LiveBench Data Analysis | 79.3% | 78.0% |
| LiveBench IF | 70.0% | 60.1% |
| LiveBench Language | 81.2% | 72.6% |
| LiveBench Reasoning | 86.7% | 85.6% |
| SimpleQA Verified | 42.9% | 41.0% |
| Maths | ||
| FrontierMath (tiers 1-3) | 67.4% | 82.1% |
| AIME-style maths (OTIS mock) | 98.3% | 98.3% |
| LiveBench Mathematics | 93.3% | 87.2% |
| ProofBench v1.1 | 54.0% | 60.0% |
| Chess puzzles | 48.0% | 43.0% |
| FrontierMath (tier 4) | 26.8% | 61.0% |
| Surface Evolver Bench | 46.3% | 61.9% |
| Coding | ||
| Terminal-Bench 4.0 | 19.7% | 11.6% |
| LiveBench Agentic Coding | 77.3% | 48.4% |
| LiveBench Coding | 80.0% | 82.9% |
| Code Migration | 45.6% | 44.5% |
| CyberBench v1.1 | 73.7% | 73.6% |
| IOI | 40.3% | 61.8% |
| ProgramBench | 0.5% | 0.0% |
| Vibe Code Bench 1-100 | 16.4% | 22.6% |
| Vibe Code Bench v1.1 | 84.7% | 77.1% |
| Agents | ||
| APEX-Agents | 39.5% | 43.0% |
| BioMysteryBench | 67.8% | 61.5% |
| SkillsBench | 69.8% | 60.5% |
| SRE Bench | 0.8% | 1.5% |
| Terminal-Bench-Science | 0.0% | 0.0% |
| DTBench | 89.9% | 89.1% |
| Furniture assembly | 34.2% | 42.5% |
| LMCA | 47.0% | 48.5% |
| Mystery game puzzles | 43.0% | 21.0% |
| Professional work | ||
| Finance Agent (v2) | 53.5% | 55.0% |
| EMB | 57.2% | 67.1% |
| LegalBench | 83.3% | 84.0% |
| Legal Research Bench | 41.3% | 36.5% |
| MedCode | 41.2% | 42.4% |
| MedScribe | 85.5% | 84.4% |
| Public Benefits Bench v1.1 | 64.3% | 61.2% |
| SAGE | 47.9% | 44.2% |
| Tax Agent Bench | 26.6% | 24.8% |
| Novel problems | ||
| ARC-AGI-2 | 72.9% | 59.6% |
| ARC-AGI-1 | 94.5% | 90.7% |
| ARC-AGI-3 | 0.9% | 0.2% |
| MysteryMechanism | 21.2% | 14.4% |
Tests at least two of these models have taken. Bold rows are the core tests.
