이미지: X — 뉴스 앰프 화면 갈무리
Summary
- DeepSeek has officially released DeepSeek-V4-Pro through the app/web "Expert Mode" and via API
- Users can set reasoning intensity to low, high, or max, and the model offers one-click support for the OpenAI Responses API for Codex
- It scored 87.9 on Terminal Bench 2.1, surpassing Opus-4.8 (85.0) but narrowly trailing Kimi-K3 (88.3)
- 모델명
- DeepSeek-V4-Pro-0813 (GA)
- 공개일
- 2026년 8월 13일
- 공개 경로
- DeepSeek Chat '익스퍼트 모드', API
- 신규 기능
- 추론 강도 3단계(낮음/높음/최대) 조절, OpenAI Responses API 네이티브 지원
- Terminal Bench 2.1
- 87.9점 (Kimi-K3 88.3 / Opus-4.8 85.0)
- HLE(with tools)
- 60.0점 (Opus-4.8 57.9)
- 개발사
- 딥시크
Officially launched a day after benchmark leak
Benchmark figures that circulated on X on August 12 turned out to be accurate within a day. DeepSeek officially released DeepSeek-V4-Pro on August 13, making it immediately available through the "Expert Mode" in the DeepSeek Chat app and website, as well as via API. The benchmark numbers that had leaked earlier matched exactly what appeared in the official release materials.
What's changed
The core update is the ability to select reasoning intensity depending on the task. Both V4-Pro and V4-Flash support three levels: low, high, and max. DeepSeek recommends low for simple tasks, high for everyday agentic work, and max for complex problems. Longer reasoning improves accuracy but increases response time and cost, and this update hands that tradeoff control to the user.
Another change is native support for OpenAI's Responses API. DeepSeek said the integration is optimized for OpenAI's coding tool Codex and supports one-click setup. Existing API users can continue using the same model name, so no migration is needed—they can simply refer to the updated API documentation.
Where it stands on benchmarks
DeepSeek's comparison table includes its own models alongside GLM-5.2, Kimi-K3, Opus-4.8, and Fable 5. The results show particular strength in agentic tasks.
| Benchmark | V4-Pro-0813 | V4-Flash-0731 | GLM-5.2 | Kimi-K3 | Opus-4.8 |
|---|---|---|---|---|---|
| Terminal Bench 2.1 | 87.9 bar:88 | 82.7 bar:83 | 81.0 bar:81 | 88.3 bar:88 | 85.0 bar:85 |
| HLE (with tools) | 60.0 bar:60 | 51.5 bar:52 | 54.7 bar:55 | 56.0 bar:56 | 57.9 bar:58 |
| Cybergym | 83.3 bar:83 | 76.7 bar:77 | - | 80.0 bar:80 | 78.3 bar:78 |
| DeepSWE | 62.7 bar:63 | 54.4 bar:54 | 46.2 bar:46 | 67.5 bar:68 | 58.0 bar:58 |
Terminal Bench 2.1 measures how well an agent performs tasks in a terminal environment. V4-Pro outperformed Opus-4.8 on this benchmark as well as on Cybergym and HLE (with tools), but mostly trailed Kimi-K3 by a narrow margin.
How to use it
① The starting point is the DeepSeek Chat app or website. After logging in, selecting "Expert Mode" from the mode menu connects users to V4-Pro.
② Users reportedly can choose a reasoning intensity level—low, high, or max—before entering a query in the chat window. Low is suited for simple questions or information lookups and yields faster responses; high works for coding tasks or multi-step workflows; and max is recommended for complex math problems or tasks requiring extended reasoning.
③ Developers already using the API can continue with the existing model name and check separate documentation for how to configure the new parameters. For developers using OpenAI Codex, the one-click setup for the Responses API reduces integration steps.
④ As an example use case, setting a coding agent to high mode for sequentially executing multiple terminal commands can balance speed and accuracy. Repetitive document summarization or short Q&A tasks can use low mode to save on costs.
An extension of the price war
According to pricing revealed on August 12, Grok 4.6 is priced at $2 per million input tokens and $6 per million output tokens, while DeepSeek-V4-Pro's GA pricing is reportedly set even lower. The gap widens further compared to Claude Opus 5 ($5 input / $25 output) and GPT-5.6 Sol ($5 input / $30 output). With this official release, the open-source camp's push to combine top-tier benchmark performance with low pricing has become even more concrete.
So what does this mean
Users can now adjust cost and speed for a single model depending on task difficulty. For developers using coding agents, Codex integration is simpler, and performance on terminal-based automation tasks has risen to a level comparable with top-tier models. This official launch reaffirms the trend of pairing low prices with top-tier benchmark performance.
