DeepSeek V4 Pro: Rock-bottom Cost Per Task, But Trailing Kimi K3
Trending Topics Jakob Steinschaden ● Covered by 4 sources
DeepSeek’s new V4 Pro is out on app, web and API. It’s cheap to run, but it still trails Kimi K3 on quality.
Based on reporting by Trending Topics, Jakob Steinschaden — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
DeepSeek has put its top model into general release this week. Internally it’s called DeepSeek-V4-Pro-0813, but the company says users only need the simpler name: deepseek-v4-pro. You can use it through the app, the web, or the API.
The launch was almost suspiciously quiet. DeepSeek posted a short note and left it there, with no technical report and no long blog post to explain the changes. That’s a contrast with the usual launch routine from bigger model labs, and it fits a company that has been leaning hard on price as its main selling point.
On paper, the economics are sharp. DeepSeek lists V4 Pro at about 44 US cents per million input tokens and 87 cents per million output tokens, plus a 99 percent cache discount. Artificial Analysis calls the input side a bit pricey versus its median, but the output side sits in the middle. The more useful number is cost per completed task: around 6 cents on the Artificial Analysis Intelligence Index. That’s a little above OpenAI’s GPT-5.6 Luna at 5 cents, and far below Moonshot AI’s Kimi K3 at 84 cents. The full index run came to $135 for DeepSeek.
DeepSeek is also changing the API pricing model from 16 August, 16:00 UTC. It will switch to peak and off-peak rates, with off-peak charged at half price. That comes right after the company said usage was so strong it had to raise prices significantly and tell users to plan around it.
The catch is quality. V4 Pro scored 53 on the Artificial Analysis Intelligence Index, which is comfortably above the median for comparable models, but still behind OpenAI’s Terra by four points and Kimi K3 by seven. It also landed twelfth on the Vals AI index, behind OpenAI’s previous-generation GPT-5.5 and well behind Kimi K3 and Anthropic’s Claude Opus 5. The weak spots were clear: sandboxed terminal work, complex Excel financial models, and long coding jobs that some developers say it stops too early.
There is one area where it looks much better: cybersecurity. Belgian security firm Aikido Security says V4 Pro found more vulnerabilities than any other model it tested, including Claude Opus 5 and Alibaba’s Qwen 3.8. Even there, the win comes with a warning label: researcher Philippe Dourassov said the model also throws up a lot of false alarms. DeepSeek is also building its own agent framework, the DeepSeek Harness, and invited open-source developers to beta-test it in early August.
My take — AI-written commentary, not fact-checked reporting
This is the classic DeepSeek move: make the bill look tiny, then let someone else worry about whether the model is actually the best at the job. Cheap per task matters, and loudly so, but “good enough and inexpensive” is still not the same as leading. The market keeps rewarding models that can do fewer embarrassing things, not just models that cost less to embarrass you.
Read more about this at: Trending Topics
