How to Generate SVGs with AI Through One API

Image models have produced pixels for years. That works for photos and fails for the graphics product teams ship most: icons, logos, stickers, and spot illustrations. Those need to scale from a favicon to a billboard, change color with the theme, and live in a repository as code. Until now, the path was a PNG from a model followed by a tracing tool, which loses detail and leaves messy paths.

You can now generate SVGs with AI in one request. OpenAI's GPT Image 2.6 returns vector output natively, and LLM Gateway exposes it on the same /v1/images/generations endpoint and API key you already use. DevPass coding plans cover text models only, so image requests from them return a 403.

Generate an SVG with AI in one request

Set output_format to svg. Accounts with OpenAI SVG model selection enabled require svg.model; accounts without it must omit svg. The optional reasoning effort controls the author independently of image quality:

1curl https://api.llmgateway.io/v1/images/generations \2  -H "Authorization: Bearer $LLM_GATEWAY_API_KEY" \3  -H "Content-Type: application/json" \4  -d '{5    "model": "openai/gpt-image-2.6-flare",6    "prompt": "A minimal line icon of a paper airplane, single stroke color, no background",7    "output_format": "svg",8    "svg": { "model": "gpt-5.4-mini", "reasoning": { "effort": "low" } },9    "quality": "medium"10  }' | jq -r '.data[0].b64_json' | base64 --decode > paper-airplane.svg

1curl https://api.llmgateway.io/v1/images/generations \2  -H "Authorization: Bearer $LLM_GATEWAY_API_KEY" \3  -H "Content-Type: application/json" \4  -d '{5    "model": "openai/gpt-image-2.6-flare",6    "prompt": "A minimal line icon of a paper airplane, single stroke color, no background",7    "output_format": "svg",8    "svg": { "model": "gpt-5.4-mini", "reasoning": { "effort": "low" } },9    "quality": "medium"10  }' | jq -r '.data[0].b64_json' | base64 --decode > paper-airplane.svg

The response has the usual images shape. data[0].b64_json is the base64 of a UTF-8 SVG document, and output_format echoes "svg":

1{2  "created": 1762344000,3  "data": [4    {5      "b64_json": "PHN2ZyB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciIC4uLg=="6    }7  ],8  "output_format": "svg"9}

1{2  "created": 1762344000,3  "data": [4    {5      "b64_json": "PHN2ZyB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciIC4uLg=="6    }7  ],8  "output_format": "svg"9}

The response also carries usage with the token counts and usage.cost, the billed cost of the request in USD.

The same request works from the OpenAI SDK by pointing baseURL at the gateway:

1import OpenAI from "openai";2import { writeFile } from "node:fs/promises";3
4const client = new OpenAI({5  apiKey: process.env.LLM_GATEWAY_API_KEY,6  baseURL: "https://api.llmgateway.io/v1",7});8
9const result = await client.images.generate({10  model: "openai/gpt-image-2.6-sunburst",11  prompt: "A flat bakery logo: a golden loaf above the words Field & Flour",12  output_format: "svg",13  svg: { model: "gpt-5.4-mini", reasoning: { effort: "low" } },14});15
16const svg = result.data?.[0]?.b64_json;17if (!svg) {18  throw new Error("No SVG returned");19}20await writeFile("logo.svg", Buffer.from(svg, "base64"));

1import OpenAI from "openai";2import { writeFile } from "node:fs/promises";3
4const client = new OpenAI({5  apiKey: process.env.LLM_GATEWAY_API_KEY,6  baseURL: "https://api.llmgateway.io/v1",7});8
9const result = await client.images.generate({10  model: "openai/gpt-image-2.6-sunburst",11  prompt: "A flat bakery logo: a golden loaf above the words Field & Flour",12  output_format: "svg",13  svg: { model: "gpt-5.4-mini", reasoning: { effort: "low" } },14});15
16const svg = result.data?.[0]?.b64_json;17if (!svg) {18  throw new Error("No SVG returned");19}20await writeFile("logo.svg", Buffer.from(svg, "base64"));

Sunburst or Flare

Start with Flare to find a direction, then run the final prompt on Sunburst. quality controls the reference image the SVG is drawn from, so low or medium is usually enough for flat icons.

Prompts that produce clean vectors

SVG output rewards prompts that describe a graphic, not a photo:

• Name the style. "Flat", "line icon", "two-color", "geometric", and "sticker with a white outline" lead to fewer, cleaner paths.
• Limit the palette. "Three colors: navy, coral, cream" keeps the file small and easy to retheme.
• Say what to leave out. "No background, no shadows, no gradients" avoids shapes you would delete by hand.
• Spell out the text. Logos with lettering come out best when the exact words are in quotes in the prompt.

What you get back, and the limits

The SVG is static vector content: paths, shapes, and colors, with no scripts, external resources, or embedded raster images. That makes it safe to inline in HTML or commit to a design system after your usual review.

A few rules come from the upstream API. The gateway rejects an SVG request with n above 1, with output_compression, or for a model without SVG output with a 400 before any provider is called:

• n must be 1.
• output_compression and streaming are not available for SVG.
• SVG is a generation format. /v1/images/edits returns raster images, and an SVG is not accepted as an edit input.
• Only GPT Image 2.6 models return SVG. model: "auto" routes to a raster model, so name the model explicitly.

In chat completions, set image_config.output_format to "svg", and the image arrives as a data:image/svg+xml;base64,... URL in message.images.

What AI SVG generation costs

An SVG request does two pieces of work. OpenAI renders a reference image with GPT Image 2.6, then an authoring model writes the SVG. LLM Gateway bills both from the usage OpenAI reports:

• the reference image at GPT Image 2.6 rates: $5 per million text input tokens and $30 per million image output tokens;
• the authoring model's input and output tokens at that model's catalogue rates. When those rates are unknown, authoring cost is omitted and logged for reconciliation; the reference image is still billed.

Both are included in usage.cost on the response and in the request log, so you can see what each icon cost. The models page lists current prices, and per-key usage limits stop new requests once a batch job has spent more than planned. Requests already in flight can still finish past the limit.

Why run it through LLM Gateway

• One key for every image model. Use GPT Image 2.6 for vectors, Gemini or Seedream for photos, and switch with the model field.
• Cost on every response. usage.cost and the request log show what each generation billed, including the SVG authoring tokens.

• Usage limits. Limits per organization, project, and API key stop sending new requests from a runaway loop.

• Try LLM Gateway free
• Read the SVG output docs
• See the GPT Image 2.6 changelog