Agent Atlas

Practice

The Studio

Coding agents write the software. Other models make the pictures, the voices and the video. The skill is the same: pick the right model for each part, give it a reference, and check what comes back.

One builder, one model per job

Factual
Greg Isenberg’s split: a different model for code, images, video and voice.

On 13 November 2026 Greg Isenberg posted his split (Greg Isenberg).

CodeClaude Code, Codex, Cursor, T3, Antigravity
ImagesNano Banana
VideoHeyGen and Fal
VoiceElevenLabs

A product video for under $30

Factual
Five models chained into one product demo.

The same day, Daniel Studziński said his product demo cost less than $30. He chained five models for the avatar, the shots, the video and the animations (Daniel Studziński).

  1. ChatGPT
  2. Flux
  3. Seedance
  4. ElevenLabs
  5. Claude

Hedra's own advice works the same way: make the portrait with an image model like Nano Banana 2, then animate it (Hedra; Google).

Building a website with agents

Factual
The loop the tools are built around, from brief to deploy.

The loop the tools are built around. The order is our summary; each step's facts come from the makers.

  1. Find a reference

    Start from real screens, not adjectives. Mobbin is a library of app and web screens, and agents can read it through its MCP server. (Mobbin)

  2. Design

    Claude Design, Google Stitch, Figma Make or v0. Hand v0 a screenshot and it reads the layout, colors and components. (Anthropic; Google; Figma; Vercel)

  3. Write the look down

    A design system is an instruction file for looks. Stitch moves design rules between tools in a DESIGN.md. Claude Design stores colors, type and components so later designs stay on them. (Google; Anthropic)

  4. Build

    A coding agent turns it into a site. Asked for a front end, Claude writes one HTML file it can render; v0 defaults to shadcn/ui components. (Anthropic; Vercel)

  5. Look at it

    The agent checks its work by seeing it. Claude ends a computer-use batch with a screenshot. Playwright compares each new screenshot against a saved one. (Anthropic; Playwright)

The step people skip is “Look at it”. An agent that cannot see the page will tell you it looks right. This site checks every page in a real browser, on a desktop and a phone, in light and dark, before it ships.

What a second of video costs

Estimate
Per-second prices for AI video, side by side.

Two prices are per second on the makers' pages. Runway's is our division: its $15 plan buys 52 seconds of Gen-4.5.

Hedra Avatar, 720p (Hedra)$0.05
Gemini Omni 1.2 Flash, 720p (Google)$0.07
Runway Gen-4.5, Standard plan (Runway)$0.29

A 60-second avatar video at 720p: about $3 on Hedra, $4.20 on Gemini Omni. Sora 2 is gone: OpenAI shut down its Sora 2 models and video API on 24 September 2026 (OpenAI).

What the plans cost

Factual
Monthly prices for the media tools, as printed on November 13, 2026.

Monthly prices as printed on 13 November 2026. They change often; follow the links.

ForToolPrice
WebsitesLovableFree: 5 build credits a day, up to 30 a month (Lovable)
WebsitesBoltPro $25 a month. Free: 300K tokens a day (Bolt)
Websitesv0Plus $30 per user a month. Premium $20 is closing to new users (Vercel)
DesignFigmaFull seat $16 a month with 3,000 AI credits (Figma)
ImagesMidjourney$10, $30, $60 or $120 a month (Midjourney)
ImagesFLUX 3From $0.041 an image. No subscription (Black Forest Labs)
ImagesGPT Image$8 in, $30 out per million tokens (OpenAI)
VideoKlingStandard $8.80 a month for 660 credits (Kling AI)
VideoRunwayStandard $15 a month: 52 seconds of Gen-4.5 (Runway)
AvatarsHeyGenCreator $29 a month, $24 billed yearly (HeyGen)
AvatarsCaptionsBasic $9.99 a month, 200 credits (Captions)
VoiceElevenLabsStarter $6 a month (ElevenLabs)

Say it is AI

Factual
What YouTube requires you to disclose, and whose face you may use.

YouTube requires creators to disclose realistic AI-generated or altered content, including making a real person appear to say or do something they did not (YouTube). Our advice: an avatar of yourself counts, and an avatar of anyone else needs their permission first.