An AI content factory in 33 hours, and the metric that actually matters

The system

A single Go process with a SQLite-backed queue and a React/Three.js studio. Each panel moves through layout, line art, and colour, with at most three attempts per stage and no automatic paid retries. An MCP server lets agents inspect state, claim work, and close gates, but none of its tools can generate an image.

Cost, measured vs listed

Vendor list prices for image models range from about $0.03 per image up to roughly $0.28 depending on the model. The measured spend for the first eight calls to fal was about a dollar, with warm latency of 5–12 seconds and cold starts over two minutes.

But per-image price is the wrong unit. A panel that takes four rerolls and a human rejection costs five images plus review time. The project’s own handoff notes define the right metric, dollars per accepted panel, never dollars per image, and the pipeline doesn’t track it yet. That gap is the most useful finding.

Context is a cost too

One full read of the factory state costs an agent about 30,000 tokens. An agent told to “inspect before every write” will pay that on every action. Tool design is cost design.

What I’d tell a team running generative pipelines

Put gates where work is claimed, keep spend out of the agent’s tools, repair instead of regenerate, and count the rejects. The invoice will look cheap right up until you divide by what you actually shipped.