Model Details
DeepSeek V4 Pro
Quick answer: DeepSeek V4 Pro is DeepSeek's flagship model in the V4 generation, featuring a 1M-token context window, up to 384K max output, and support for both thinking and non-thinking reasoning modes. It is available via the DeepSeek API at $0.435 input / $0.87 output per 1M tokens, making it one of the most cost-effective frontier models available. On LHTB it scored 0.307 mean reward (rank 12), solving 3 of 46 long-horizon terminal tasks.
At a Glance
Where DeepSeek V4 Pro leads
- Exceptional price-performance ratio — $0.435/$0.87 per 1M tokens at frontier capability.
- 1M context window with 384K max output — among the most generous output limits available.
- Flexible thinking mode: switching between fast non-thinking responses and deep reasoning on demand.
Where it lags
- Not yet widely available via international cloud providers; primarily through DeepSeek's own API.
- LHTB score of 0.307 places it below the Anthropic and OpenAI flagship cluster.
Best for: Cost-sensitive agentic pipelines, long-document reasoning, and high-throughput API workloads requiring frontier-level intelligence.
What DeepSeek V4 Pro Is
DeepSeek V4 Pro is the higher-capability tier of DeepSeek's V4 model family. It succeeded the V3 line and introduced substantially stronger Agent capabilities and top-tier reasoning alongside the more economical V4 Flash variant. The Pro tier is positioned as the flagship option when full reasoning power and maximum context handling are required.
The model supports both thinking and non-thinking modes. In thinking mode, it engages an extended chain-of-thought process for tasks that benefit from deliberate reasoning — math, science, and complex planning. In non-thinking mode, it returns faster responses at the same price per token, making it suitable for lower-latency pipelines without sacrificing model capability. Both modes share the same 1M-token context window and 384K output limit.
DeepSeek V4 Pro's pricing is notable: at $0.435 per 1M input tokens and $0.87 per 1M output tokens (cache miss), it delivers flagship reasoning at a fraction of the cost of comparable Anthropic or OpenAI models. On LHTB — a hard 90-minute-per-task terminal agent benchmark — it scored 0.307 mean reward and solved 3 of 46 tasks, placing 12th overall.
Specifications
| Field | Value |
|---|---|
| Context window | 1,000,000 tokens |
| Max output | 384,000 tokens |
| Modalities | Text |
| Reasoning modes | Thinking + Non-thinking |
| Input price (cache miss) | $0.435 / 1M tokens |
| Input price (cache hit) | $0.003625 / 1M tokens |
| Output price | $0.87 / 1M tokens |
| License | Proprietary (API only) |
| API model name | deepseek-v4-pro |
Related Benchmarks
- Long-Horizon Terminal-Bench (LHTB) — 0.307 mean reward, rank 12 (3/46 solved)
- SWE-bench Verified — industry standard for software engineering agents
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.