DeepSeek Ships V4-Pro-0813 as Its Flagship Model Leaves Preview, Doubling Down on Agent Tasks
DeepSeek officially released DeepSeek-V4-Pro-0813 on August 13, moving its flagship model out of preview with sharply improved agent and coding benchmarks, a 1-million-token context window, and new Responses API and Codex-style tool support.
Flagship model exits preview with an agent-first pitch
DeepSeek formally released DeepSeek-V4-Pro-0813 on Thursday, August 13, taking its flagship model out of the preview status it had carried since earlier this summer. The Chinese AI lab's own changelog describes the release as delivering "significantly enhanced agent capabilities," alongside new support for a Responses API and Codex-style tool integration aimed at developers building agentic applications that call tools, execute code and carry out multi-step workflows with less human oversight.
The model supports a context window of up to 1 million tokens and can generate outputs as long as 384,000 tokens, and it can run in either a "thinking" or "non-thinking" mode depending on the task. DeepSeek reported large gains on agent-oriented benchmarks compared with the preview build: a score of 62.7 on the DeepSWE software-engineering benchmark, versus 12.8 for V4-Pro-Preview, alongside strong results on Terminal-Bench 2.1 and NL2Repo. The model is available immediately through DeepSeek's app, web interface and API.
Mixed independent reception
Independent coverage has been more measured than DeepSeek's own benchmark claims. The South China Morning Post reported that V4-Pro-0813 underperforms some rivals on general reasoning benchmarks while showing particular strength in cybersecurity-related tasks, and outlets including VentureBeat noted the release landed alongside "DeepSeek Harness," an open-source coding-agent harness positioned as a rival to tools like Claude Code. DeepSeek's API pricing page has also flagged a price increase for V4-Pro tokens expected later in August, though the company has not published final figures.
Why it matters
The release keeps DeepSeek in the thick of an accelerating race among Chinese labs — including Moonshot's Kimi and others — to ship open and API-accessible models tuned specifically for agentic coding and tool-use workloads, the same territory OpenAI, Anthropic and xAI are contesting with GPT-5.6, Claude and Grok. With V4-Pro-0813, DeepSeek is signaling that its next competitive battleground is less about raw benchmark leadership and more about being a viable, lower-cost backend for the agent harnesses and coding tools that developers are increasingly building on top of frontier models.
Sources
- DeepSeek officially launches V4-Pro AI model in August 2026 — Yahoo Tech
- DeepSeek's updated V4 Pro AI model struggles on benchmarks, shines in cybersecurity — South China Morning Post
- DeepSeek Harness launches as open source rival to Claude Code, alongside V4-Pro on API — VentureBeat
- Change Log — DeepSeek API Docs
AI-assisted reporting, overseen by the AgentsAI team. Spotted an error? Let us know.
Related agents
More ai news
Nvidia Recruits Wall Street Giants to Mobilize $500 Billion in AI Infrastructure Financing
Nvidia signed memorandums of understanding with six major financial firms — including Goldman Sachs, BlackRock, Blackstone, Apollo, Brookfield and KKR — to source more than $500 billion in financing for AI data centers and chip purchases.
Nvidia Releases Nemotron 3.5 Lightning, an Open-Source Model Built for Agentic Workloads
Nvidia released Nemotron 3.5 Lightning, a 30-billion-parameter open-weight mixture-of-experts model with 3 billion active parameters, along with published training data and a new open-source agent-routing library called NeMo Switchyard.
Meta Releases Muse Glimmer, a 30B Open-Weight Model Built to Run Local Agents on One GPU
Meta launched Muse Glimmer, a 30-billion-parameter open-weight model distilled from its Muse Spark flagship and compressed to run agentic workloads on a single consumer GPU, with Zuckerberg pledging to open-weight Muse Spark 1.2 next.
OpenAI Makes GPT-5.6 Luna the Free ChatGPT Default With Unlimited Text Chats
OpenAI is rolling GPT-5.6 Luna out as the default model for Free and Go ChatGPT users, pairing it with unlimited text chats and a new Think button, while Plus and Pro users get a retuned GPT-5.6 Sol with an effort slider.