DeepSeek V4 plugs into OpenClaw

DeepSeek’s latest model is beginning to look less like a routine upgrade and more like a calculated push to reset the rules of the AI industry.

The signs began to emerge from its launch last week, when it was revealed that with its latest V4 model, DeepSeek has shifted away from Nvidia hardware.

The Chinese artificial intelligence startup has optimized V4 to run on Huawei’s Ascend chips, a move that reflects both technical necessity and geopolitical intent, in light of the tightened US export controls that have accelerated efforts within China to build a self-reliant AI stack.

Now, days after its release, DeepSeek V4 has already been adopted by OpenClaw, which made the lighter V4 Flash its default model while adding the more powerful V4 Pro to its platform. The move places the Chinese startup’s technology directly within one of the emerging generation of AI agents that has taken the world by storm and is particularly popular in China.

A three-pronged push

As noted in coverage of the release, besides OpenClaw, the DeepSeek V4 models are optimized for other mainstream agents as well, including Anthropic’s Claude Code and Tencent Holdings’ CodeBuddy.

Notably, all these AI agents consume large volumes of tokens. DeepSeek is also looking to capitalize on that aspect by undercutting rivals aggressively.

At launch, DeepSeek’s V4 Pro model was already cheaper than its rivals, priced at $3.48 for 1 million tokens of output. For comparison, the same amount of tokens costs $30 and $25 for OpenAI and Anthropic, respectively. On Monday, it went a step further by announcing a further 75% promotional discount. Under this offer, which runs until May 5th, 2026, a million V4-Pro input tokens cost a mere $0.036.

Taken together, the rollout reveals DeepSeek’s three-part strategy. Cheaper tokens make agent deployment viable at scale, tighter integration with popular agents drives usage, and hardware alignment reduces external dependencies.

At launch, DeepSeek said the V4-Pro version outperforms other open-source models on world-knowledge benchmarks, while acknowledging that it trails Google's closed-source Gemini-Pro-3.1 model.

However, it appears that DeepSeek is playing a longer game by focusing less on headline performance and more on controlling how AI is built, deployed, and priced.