V4 Pro
DeepSeek · released Apr 24, 2026
DeepSeek V4 Pro is a large-scale Mixture-of-Experts model from DeepSeek with 1.6T total parameters and 49B activated parameters, supporting a 1M-token context window. It is designed for advanced reasoning, coding,...
Specification
- Context window
- 1.05M
- Max output
- 384K
- Knowledge cutoff
- Not stated
- Parameters
- 1598.8B
- Licence
- mit
- Serving providers
- 18
- Moderated
- No
- Uptime
- 100.0%
Intelligence
45.3
93th percentile
Coding
59.4
Coding Index
Agentic
37.8
Agentic Index
Output speed
71 t/s
Median across providers
Latency
380ms
Time to first token
Cost per task
$0.09
Estimated
Benchmarks
Where the score comes from
The Intelligence Index is a composite. These are the underlying evaluations this model was actually measured on.
Evaluation scores
Percentage correct · higher is better
- τ²-bench (Telecom)96.2%
- GPQA Diamond88.8%
- IFBench76.5%
- AA-LCR (long context)70.0%
- SciCode50.0%
- Terminal-Bench Hard46.2%
- Humanity's Last Exam37.5%
An evaluation missing from this list was not run for this model — it is not a zero.
View as table
| Evaluation | Score |
|---|---|
| τ²-bench (Telecom) | 96.2% |
| GPQA Diamond | 88.8% |
| IFBench | 76.5% |
| AA-LCR (long context) | 70.0% |
| SciCode | 50.0% |
| Terminal-Bench Hard | 46.2% |
| Humanity's Last Exam | 37.5% |
Against its peers
Intelligence Index · this model highlighted, nearest peers in grey
- Qwen3.7 MaxQwen46.7
- GPT-5.3-CodexOpenAI45.5
- MiniMax M3MiniMax45.4
- V4 ProDeepSeek45.3
- Kimi K2.6Moonshot AI45.1
- GPT-5.2OpenAI43.3
- Kimi K2.7 CodeMoonshot AI43.0
- MiMo-V2.5-ProXiaomi42.9
Peers are the models sitting closest on the Intelligence Index — the set you would realistically choose between.
View as table
| Model | Intelligence |
|---|---|
| Qwen3.7 Max | 46.7 |
| GPT-5.3-Codex | 45.5 |
| MiniMax M3 | 45.4 |
| V4 Pro | 45.3 |
| Kimi K2.6 | 45.1 |
| GPT-5.2 | 43.3 |
| Kimi K2.7 Code | 43.0 |
| MiMo-V2.5-Pro | 42.9 |
Percentile among all indexed models
Pricing
What it costs to run
List prices per million tokens, plus what one representative task works out to.
List price
- Input / 1M tokens
- $0.435
- Output / 1M tokens
- $0.87
- Cached input / 1M
- $0.004
- Blended 3:1
- $0.544
One task, estimated
$0.09
- Input tokens
- 50,000
- Output tokens
- 80,000
- Profile
- Reasoning
Estimated from list pricing: 50K input tokens plus 80K output tokens for reasoning models (25K for non-reasoning).
Serving providers
- DeepSeek
- StreamLake
- GMICloud
- DigitalOcean
- Ionstream
- Novita
- Cloudflare
- Baidu
- DeepInfra
- Alibaba
- SiliconFlow
- Venice
- AtlasCloud
- BaseTen
- Parasail
- Together
- CoreWeave
- Fireworks
Speed and latency figures are medians across these providers, so a widely-served model reports a blend rather than any single endpoint.
Arena
Head-to-head generation quality
Elo from pairwise judgements, broken out by the kind of thing the model was asked to build.
- Overall Elo
- 1,313
- Win rate
- 58.6%
- Strongest at
- 3D scenes
- Tournaments
- 3,541
Elo by category
Dot position on a 1,175–1,325 scale · Elo has no meaningful zero
- 1,313
- 1,274
- 1,259
- 1,257
- 1,226
- 1,182
Agent categories (full-stack apps, mobile apps) are only scored for models tested in agent harnesses.
View as table
| Category | Elo | Win rate |
|---|---|---|
| 3D scenes | 1313 | 58.6% |
| Game development | 1274 | 55.2% |
| Websites | 1259 | 52.2% |
| UI components | 1257 | 51.6% |
| Data visualisation | 1226 | 49.6% |
| SVG | 1182 | 46.4% |