DeepSeek · DeepSeek V4
DeepSeek V4 Pro (0813)
DeepSeek's larger open-weight text MoE (1.6T total / 49B active, 1M context), GA since the 0813 post-training update, for harder coding and agent work.
- AA Intelligence
- 36#29 of 142
- Input / output per 1M
- $1.32/ $3.96
- Output speed
- 89tok/s
- Context window
- 1M tokens
Composite indexes
- Artificial Analysis Intelligence Index#29 of 14236
Leader: Claude Opus 5.5 at 58
Coding
- Terminal-Bench 4.0#24 of 8814.1%12.4%
Leader: Claude Mythos 5.1 at 60.9%
- Terminal-Bench 2.x#26 of 10387.9%78.7%
Leader: Claude Fable 5.1 at 91.4%
- SciCode#36 of 9451%
Leader: Claude Opus 5.5 at 66.9%
Agents and tool use
- GDPval-AA#24 of 1031441
Leader: Claude Opus 5.5 at 1846
Reasoning and knowledge
- GPQA Diamond#18 of 14292.8%92.4%
Leader: GPT-6 Astra at 96.1%
- Humanity’s Last Exam#32 of 14760%42.7%41%
Leader: Claude Opus 5.5 at 61.4%
- CritPt#20 of 13918%
Leader: GPT-5.6 Sol at 32.3%
Vision and long context
- AA Long Context Reasoning#27 of 13980.3%
Leader: Kimi K3 at 88.7%
Human preference
- LMArena Text#24 of 541463
Leader: Claude Fable 5 at 1506
- LMArena WebDev#24 of 541581
Leader: Claude Opus 5.5 at 1818
Other published results
| Result | Score | Reported by | Source |
|---|---|---|---|
| DeepSWE | 62.7% | vendor | DeepSeek-V4-Pro-0813 model card (Hugging Face) ↗ |
| Agents' Last Exam | 25.7% | vendor | DeepSeek-V4-Pro-0813 model card (Hugging Face) ↗ |
| Toolathlon-Verified | 74.1% | vendor | DeepSeek-V4-Pro-0813 model card (Hugging Face) ↗ |
| NL2Repo | 61.5 score | vendor | DeepSeek-V4-Pro-0813 model card (Hugging Face) ↗ |
| CyberGym | 83.3% | vendor | DeepSeek-V4-Pro-0813 model card (Hugging Face) ↗ |
| AutomationBench (Public) | 31.8% | vendor | DeepSeek-V4-Pro-0813 model card (Hugging Face) ↗ |