gpt-image-2.6-all SVG output falls short — SVG is better served by the “write code” route
A user reported that SVG from -all is not usable; our retest confirms it. output_format=svg does return real vectors, but no writer model is involved.
The cause is the route. gpt-image-2.6-all is the reverse-engineered ChatGPT web pipeline: SVG comes straight from the image model, and the svg.model field is accepted but ignored (documented on the model page), so you cannot point it at a stronger writing model. Official-relay gpt-image-2.6-sunburst / gpt-image-2.6-flare take a different path: svg.model (for example gpt-5.6-terra) writes the SVG code first and returns it as a vector — code is the image, with much better control over structure and text.
Options, by use case:
- Official relay
gpt-image-2.6-sunburst/gpt-image-2.6-flare: choose the writing model withsvg.model, tunesvg.reasoning.effort, billed by tokens - A text model writes the SVG:
gpt-5.6-terra/gpt-6-lunaemit<svg>code, fully controllable gpt-image-2.6-all: still returns SVG, but thesvgobject has no effect and quality falls short; good for a quick preview
← Back to Live Updates · 📚 Monthly archive