Original video in Chinese.
Key Takeaway
- DeepSeek’s new V3 0324 model performs exceptionally well in MCP calls, offers outstanding value for money, and comes close to Claude-3.7 Sonnet in performance.
- The DeepSeek model has clear reasoning and planning abilities. It can break user needs into explicit tasks and determine which tools are needed.
- DeepSeek’s tool-calling ability has improved significantly, and together with its cost advantage, it will drive further adoption of MCP.
- The article predicts that the evolution of Agents will shift from task orchestration to letting the model play freely, with a super-strong model at the core and a huge number of atomic tools alongside it.
- DeepSeek’s progress, together with the MCP protocol, signals that the AI industry is about to enter a period of rapid growth.
DeepSeek’s newly released V3 0324 model may be the most cost-effective model for running MCP. In terms of performance, it’s close to Claude-3.7 Sonnet. It calls MCP very smoothly, but the cost is much, much lower — honestly, it’s ridiculously cheap. I’ll show you and you’ll understand right away.
This is Cline, the MCP client I use the most right now. It already supports DeepSeek’s latest model. I’m using the paid version here. The platform does offer a free version, but I don’t recommend it. I tried it earlier, and it was way too slow; once the steps got a little more complex, it would easily get interrupted. It was pretty annoying. So let’s just pay for it honestly.
My need is very simple: Google released the Gemini 2.5 model. This is their official blog. I asked DeepSeek to scrape the content, translate it into Chinese, add a summary at the beginning, and then save it into a document.
You can see that DeepSeek first made a four-step plan:
First, it broke the user’s request into clear tasks;
Second, it determined which tools to use, including the extraction tool in Tavily MCP and the tool for writing to a file;
Third, the current environment is that the file already exists, and the user has also allowed direct operation;
Fourth, it set execution steps for itself.
This is where AI is stronger than humans. Think about it — how many people can think and plan this clearly?
The whole process took two or three minutes, so I won’t show it in detail. After the document was written, the whole task ended, and the total cost was $0.0358.
Next, let’s make it more difficult. I asked it to call two MCPs: one was Sequential-thinking, with as many steps as possible. The other was Tavily, responsible for searching for information online. Before each step of thinking, it had to search for information once, then think based on what it found.
A request like this is a real test for a model. It has to know how to break down the problem, adjust its line of thinking at any time based on the information it actually finds, decide what to search for next, and frequently call MCPs without making mistakes.
I recommend that after watching the video, you test it this way too, whether you’re testing the model or the client. Then you’ll know how to choose.
Back to DeepSeek. The whole reasoning process took just over three minutes. DeepSeek did six rounds of thinking in total and finally gave the answer. But I felt the key points below weren’t detailed enough, so I asked it to improve them further. In the end, DeepSeek completed the answer for $0.039.
From these two examples, you can see that DeepSeek’s new model has no problem using MCP, and it’s very cheap. To be honest, I’ve been using Claude to run MCP for this period, and I’ve already spent more than ten dollars on API costs. If you use it frequently in daily life, it really hurts.
High cost-effectiveness is why I recommend DeepSeek. The official account introduced this minor version upgrade in a WeChat article. The model capability improvements include several aspects, such as better performance on reasoning tasks, stronger front-end development ability, upgraded Chinese writing, and so on.
What I value most, and what I think is the most important, is the improvement in tool-calling ability. It’s still the logic I mentioned before:
There are two paths for AI development: one is to obtain more information, and the other is to call more tools.
If it can only process text and can’t do multimodal work, then AI’s world is black and white. That’s why I’m optimistic about Gemini.
If it’s limited to reasoning and can’t use more tools, then AI only has a brain but no hands or feet. That’s why I’m optimistic about Claude.
Now DeepSeek has finally improved its tool-calling ability, and combined with its already strong cost advantage, it will definitely push MCP toward broader adoption.
Finally, let me share my judgment on Agent development this year:
The task-orchestration model represented by Dify and Coze will gradually be replaced. Although this approach is highly accurate, it consumes too much human labor, has a high barrier to entry, and severely limits what the model can do, so it is only suitable for enterprise production environments.
I believe the optimal solution must be: humans only need to set the starting point — that is, the Context, the context — and the endpoint, that is, the goal. Everything between the starting point and the endpoint is left for the model to play out freely.
A super-strong model as the single core, paired with a massive number of atomic tools, is the focus of the AI industry this year.
Now, we already have models like Claude and DeepSeek, and we also have middleware protocols like MCP. Everything is in place, and everything will unfold at high speed.
OK, that’s all for this episode. If you want to learn about AI, come to our newtype community. See you next time!