Best cell per row highlighted. Null means undisclosed — never counted as zero.
| Provider | DeepSeek | DeepSeek |
| Family / tier | DeepSeek V3 · Large | DeepSeek V4 · Flash |
| Status | GA | GA |
| Released | 2025-08-21 | 2026-09-10 |
| Knowledge cutoff | 2025-06 | — |
| Context window | 128K | 1M |
| Max output tokens | 8K | 384K |
| Input modalities | text | text, image |
| Reasoning mode | optional | optional |
| Open weights | Yes | Yes |
| Input | $0.21 | $0.3 |
| Output | $0.79 | $1.2 |
| Cached input | $0.021 | $0.006 |
| Batch input | — | — |
| Free tier | No | Yes |
| Output speed (tok/s) | — | — |
| Time to first token (s) | — | — |
| Latency tier | medium | fast |
| Trains on inputs | Yes | Yes |
| License | MIT | MIT |
| Clouds | — | — |
| Compliance | — | — |
| Unified score | 7.7 | 8.6 |
| Decision Maker | 7.5 | 8.5 |
| Domain Strategist | 7.5 | 9 |
| Finance Lead | 8.5 | 9 |
| Domain Practitioner | 8 | 9 |
| Power User | 7.5 | 8.5 |
| Skeptic | 7.5 | 7.5 |
| Value score | 8.8 | 9.6 |
| MMLU-Pro | 83.7 | — |
| GPQA Diamond | 74 | 90.9 |
| AIME 2025 | 88 | — |
| SWE-bench Verified | 66 | — |
| LiveCodeBench | 80 | — |
| Aider Polyglot | 76.3 | — |
| Terminal-Bench | — | 90.6 |
| Humanity's Last Exam | — | 36.8 |
| LMArena Elo | 1417 | — |
| LMArena Coding Elo | 1461 | — |
Scores link to their sources. Missing cells mean the vendor hasn't disclosed a result — honesty over padding.