Technokafe Built a Video Essay With Code: 73 Seconds for $10 in Claude Code Tokens

The author of the Telegram channel Technokafe has so far made one-minute reels. Now he is trying 16:9 video essays, where a single topic is broken down into chapters. The first chapter runs 73 seconds and is about one-person companies. There is no editing software in the process: the script, shots, voice and scenes are described in text, and Claude Code on Opus 5.5 turns that description into a video. By the author's estimate, the agent's work cost about $10 at API rates, or roughly 1% of the weekly limit on a Max subscription.

The number the chapter rests on

The chapter is built on data from the Solo Founders Report, which Carta published in December 2025. In the US, the share of new startups with a single founder rose from 23.7% in 2019 to 36.3% in the first half of 2025. That is an increase of 12.6 percentage points, or growth of about 1.5x in six years. For a channel about vibe coding and AI agents, the thesis fits well: the figure describes the audience that builds products alone with the help of agents.

The author himself appears on camera. He explains how he splits work between Claude Code, Codex, n8n and Make, shows the Carta report with the key figure highlighted, and walks through an example on a cork board. The logic of the format is simple: a reel carries one idea, while an essay has room for an example, a sourced figure and a conclusion.

What the video is made of and what it cost

The author listed the entire stack with prices:

• Claude Code (Opus 5.5) — the orchestrator. It receives a text description of the script, shots, voice and scenes and assembles the video from it. About $10 via API, roughly 1% of the weekly Max limit.
• ElevenLabs (voice) — a clone of the author's voice, $0.25.
• ElevenLabs (music) — the soundtrack is generated in the same service; the price is not given separately.
• Veo 3.1 Lite via Vertex — animates the author's photo and turns it into a presenter. $0.80 from Google credit.
• Volcengine via kie — lip-sync, only on segments with a face. About $0.50.
• GPT Image via Codex — shot generation. It runs on a ChatGPT subscription, with no separate charge for images.
• HyperFrames — editing in code instead of a traditional editor.
• SAM 2.1 — cuts the speaker out of the frame locally, on the author's GPU.
• n8n and Make — not part of the video's pipeline; they appear in the chapter itself as part of the author's workflow, alongside Claude Code and Codex.

Adding up the named paid items ($10 + $0.25 + $0.80 + $0.50) gives about $11.55 for 73 seconds of finished video, roughly 16 cents per second. The ChatGPT and Max subscriptions are not included in that sum, and part of the costs was covered by Google credit.

Why modularity matters

The main engineering idea is that each chapter lives on its own. A long video is made in parts, and one chapter can be redone without touching the others. It is the same logic as in software development: small, independent modules that are easy to rebuild. The agent works from text descriptions, so a fix comes down to changing text rather than re-editing a timeline.

Cost scales linearly. A ten-minute essay is about eight chapters of 73 seconds each. At the same per-chapter cost, that comes to roughly $90–95 per video, not counting subscriptions. This is a rough estimate based on the author's figures: actual costs depend on the number of segments with a face and the number of redos.

What you can take from this

Repeating the approach does not require the full stack. The minimum set is an agent for assembly (Claude Code), an image generator (GPT Image via Codex) and code-based editing (HyperFrames). Voice, a presenter and lip-sync can be added as needed. The most expensive operations are the agent's own work and video generation with a face, so that is where it pays to save: split the script into independent chapters and apply lip-sync only to segments where the person is seen up close. If you test ideas on short chapters, the cost of a mistake stays within a few dollars.

The author asks his audience whether subscribers would watch such essays at 10 or 15 minutes, and invites them to suggest a topic for the first episode.