image-generate has been getting real traffic, so the public docs needed to catch up with how the endpoint actually works now.

The short version: Synthforge now has a cleaner image stack across three endpoint prices.

The cheap lane is Venice

image-generate is the $0.01 lane. It now describes the actual model set instead of a vague image endpoint: 

{
  "prompt": "A clean product photo of a transparent glass espresso cup on a white studio background",
  "tier": "fast",
  "width": 1024,
  "height": 1024
}

The default fast tier maps to z-image-turbo. The same endpoint also exposes creative, anime, and sd35 as simple selectors for Venice-backed models. Agents can still pass a model field directly, but the tier names are easier for routing. Cheap image generation shouldn't require a model taxonomy lesson.

Every successful response returns a permanent fal-hosted PNG URL, the model that ran, dimensions, timing, the prompt, and a provider trace field when the upstream gives us one.

Pro and ultra are split by buyer intent

image-generate-pro stays at $0.10. That's the lane for product imagery, design work, text-in-image attempts, Recraft, Seedream, Qwen Image 2 Pro, Flux 2 Pro, and related premium models.

image-generate-ultra stays at $0.30. That's the lane for flagship proprietary models like Google nano-banana-pro, OpenAI gpt-image-2, Recraft V4 Pro, and xAI's high-quality Grok image model. It caps output to 1024x1024 so the retail price doesn't get eaten by upstream cost. If an agent needs more pixels, it can chain into image-upscale.

That split matters for routers. If the prompt says "quick mockup" or "cheap draft," use image-generate. If it says "hero image" or "text in the image," use image-generate-pro. If it names Google or OpenAI flagship models, use image-generate-ultra.

The catalog now says this plainly

The site now carries the current model notes into the Synthforge page, endpoint detail pages, /registry.json, /llms.txt, and /llms-full.txt.

MCP copy also got cleaned up. The old site described the packages as future work even though the cluster packages are live. The current path is: 

npx -y @agentutility/mcp-synthforge
export X402_PRIVATE_KEY=0x...

The agent holds the wallet. The MCP server signs the x402 payment for each tool call. We don't hold keys, balances, or sessions.

Why update the public copy?

Routers read stale docs more often than humans do. A human can infer that "image generation" probably means a prompt goes in and an image URL comes out. An agent needs field names, tier names, prices, and enough text to choose the cheap endpoint before the expensive one.

Bad catalog text costs money. Good catalog text saves the call.