DeepSeek Builds 3D Web Page for About 4 Yuan; Closed Models 100+

A viral video shows a Tesla accelerating from the ground, gripping and climbing a slope, flying off upside down at the crest, then hovering in mid-air. Ifanr fed this video, along with several widely shared prompts, to the new DeepSeek model, V4.1 Flash 1017. After 35 minutes, a directly runnable, animated 3D web page was delivered: the runway and car were built, and the acceleration, climb and hover flowed together. The article says it had “almost no bugs,” and the whole process cost about 4 yuan.

The challenge was that a single output had to do several things at once: build the scene, link the acceleration, climb and hover together, and deliver a complete, working project. The model did not paint the images; it wrote the web code. The article also notes that its command of Three.js, a three-dimensional web code library, means it can even be used to make games. This version's background is as follows: in September, DeepSeek updated V4.1 Flash; this Saturday, version 1017 was released, improving performance while shortening task run times and output length.

Here is how the bill adds up. The project consumed 37.52 million tokens (the unit large models use to measure text processing), of which 330,000 were output tokens, with a cache hit rate as high as 99.6%. At peak hours, cached input costs 0.04 yuan per million tokens and output costs 8 yuan per million tokens; at these unit prices, the total comes to about 4 yuan. For comparable results, the article says the same job would need at least 100 yuan on GPT-6.1 Sol and Claude models. Using a ChatGPT Plus subscription on Codex, it says, could exhaust the five-hour quota, and only the GPT-6.1 Max model at its deepest reasoning setting was available.

But these are Ifanr's own tests, based on only a handful of cases. The Tesla recreation is “not quite 1:1,” and the Eiffel Tower “doesn't look quite right.” There are many details, yet they still need checking against reference material for proportion and structure. A Markdown editor recreation also had a gap: it required the user to press Enter before rendering, a problem the article says was quickly fixed. The benchmark figures in the article come from the cited rankings; the article offers no independent retest.

Who benefits? The first to gain are people with one-off small needs: adding an interactive demo to an article, or setting up a temporary page for an event. These tasks may not justify hiring a developer, but spending a few yuan to produce a first version lowers the barrier to experimenting considerably. For closed-source model vendors, the pressure comes from the price gap. As more tasks can be handled by open-source models with results close to or even better than theirs, they must answer how far ahead they are, how long that lead will last, and whether users should pay more. The article argues that capabilities that today require expensive subscriptions may become standard in cheaper models within time, though it gives no timetable.