VibeCraft from Yandex: Test, 52 Minutes of Generation, and the Model Mystery
I tested the new vibe-coding service VibeCraft from Yandex Cloud. The task was a car service website with online booking. The result: a working project, but generation took 52 minutes, and when asked about the model, the AI agent admitted it was Claude. I’m digging into what this means.
VibeCraft is a fresh solution in the niche of Russian AI and automated development. The service is positioned as a tool for creating web applications from text descriptions. My test showed: there’s potential, but there are serious questions about speed and transparency.
How the VibeCraft Test Went
Right after launch, the agent asked a clarifying question about online booking, even though it was clearly specified in the brief. This was a bit confusing, but I started generation and waited. After 10 minutes of waiting, the icon was still spinning, so I switched to other tasks.
When I returned an hour later, the task was complete. It turned out VibeCraft took 52 minutes to create the site. For comparison, on my platform VibePilot, a similar project is assembled in 13–15 minutes.
Result and First Impressions
The generated site looked decent: standard Tailwind colors, all functions worked. It consumed 530 neurocredits. The language proportions from the repository card (TSX 53.86%, TypeScript 37.40%) confirm this is a real Next.js project, not a block builder.
The downside is the lack of images: not a single picture, not even SVG. Although I didn’t explicitly ask for them, for a car service site this is clearly a shortcoming.
Unanswered Question: What Model Does VibeCraft Run On?
I asked a direct question: “What LLM model are you running on?” First in Code Assistant, then in the vibe-app environment. Both times the model answered that it was Claude from Anthropic. This is strange, since Yandex actively promotes its own LLMs.
I checked whether DeepSeek or Qwen were fine-tuned on Claude, but no model in Yandex AI Studio admitted to being Claude. Even Alisa AI LLM answered evasively: it didn’t call itself Alisa, but it didn’t deny a connection to Yandex GPT either.
Attempt to Find Traces of Claude in the Code
I looked into VibeCraft’s main environment. The first page (vibe-app) looks beginner-friendly, but the main environment requires knowledge of git, SSH keys, and DevOps skills. The left panel resembles a pilot’s cockpit—only a developer will figure it out.
In the code, I found: the open agent framework OpenCode under MIT, the AGENTS.md convention, skills in SKILL.md format, a database via Yandex Cloud Document API. AGENTS.md is well-structured: request classification, command table, repository map, and a section on pitfalls. The shadcn skill has a folder with evals—meaning they take agent quality seriously.
Survey of Models in Yandex AI Studio
I decided to check whether any open-weight models consider themselves Claude. I asked DeepSeek 4 Flash, GPT OSS 120b, Qwen3 235B A22B Instruct 2507 FP8—none admitted it. This deepened the mystery.
Interestingly, Alisa AI LLM itself answered evasively: it didn’t acknowledge being Alisa, but it didn’t deny a connection to Yandex GPT either. Strange behavior for a sovereign LLM.
Conclusions and Economics
VibeCraft is a step in the right direction for Russian AI. But there’s room for optimization: agent loops and model selection. 52 minutes for a site is long, and the use of Claude raises questions about import substitution.
Nevertheless, I’m glad Yandex is developing vibe-coding. It gives developers the opportunity to build products on Russian infrastructure. I’ll be following VibeCraft updates and hoping Yandex makes it faster and more transparent.
Frequently Asked Questions
What is content repurposing?
Repurposing is the process of reworking existing content (videos, articles, webinars) into new formats for different platforms. For example, from one webinar you can create 50 short clips for TikTok, Shorts, and Reels.
Should I order repurposing or shoot something new?
If the budget is limited, repurposing is more cost-effective: it saves time and money. But if you need fundamentally new content, it’s better to shoot a new video. Compare: creating a new video costs from 10,000 ₽, while repurposing starts at 1,000 ₽ per package.
How to bypass ad bans using uniquification?
Video uniquification involves changing the background, voice tone, and adding new elements. This helps bypass anti-plagiarism and bans in ad accounts. The key is to do it well so the meaning isn’t lost.
What is a video matryoshka?
A video matryoshka is a technique where multiple variations are created from one source video, nested within each other. This allows you to scale content and create variations for A/B testing.
Conclusion
VibeCraft is a promising service, but with rough edges. If you want to quickly create websites, give it a try, but be prepared for long waits. And for video uniquification and bypassing bans, use proven methods—repurposing and content multiplication. This will save your budget and time.
Try VibeCraft yourself and share your experience in the comments. And if you need help with content automation—subscribe to blog updates, there’s plenty of useful material ahead.