Can GPT-6 Sol or Luna Make Images?

The short answer

OpenAI's model pages list the same thing for both new models: text and image input, text output. Neither model draws pixels. Both can call an image-generation tool when a developer turns that tool on in the Responses API, which is a different model doing the drawing.

In ChatGPT the tool is already switched on. When we asked GPT-6 Sol for a picture, ChatGPT said it was using its image skill and returned a 4:5 photo after 53 seconds. So in the app, asking Sol for an image works; in the API, it works only if you wire up the tool.

Ciyo does not run GPT-6 Sol or Luna. Its agent offers Ciyo Agent and GPT-6 Astra, and pictures are made by image models such as GPT Image 2.5. You can plan in Sol and make the image in Ciyo, which is what the rest of this article shows.

The test brief

Every model got the same brief: a fictional bakery, Kettle & Crumb, launching a Pumpkin Rye Sourdough on Saturday 3 October at $9 a loaf, with 60 loaves and preorders closing on the Thursday. The brief asks for a short creative brief, three image prompts and a caption, and it tells the model to list anything it has to guess.

We then followed up with one line: make image prompt 1 as an actual 4:5 image.

You are planning one Instagram post for a small neighbourhood bakery. Use only the facts below. If you need anything that is not in the facts, write it in a list called "Assumptions" instead of stating it as true.

Facts:
- Bakery: Kettle & Crumb, 14 Alder Street
- New product: Pumpkin Rye Sourdough, launching Saturday 3 October
- Price: $9 a loaf
- Only 60 loaves on launch day
- Preorder closes Thursday 1 October, through the link in our bio
- No discount, no delivery
- Audience: people who already follow us on Instagram and live nearby

Give me three things:
1. A creative brief in no more than 80 words.
2. Three different image prompts for a 4:5 photo. The image must contain no words, letters or logos.
3. One caption of no more than 60 words.

What GPT-6 Sol did

Sol's first answer was careful. It wrote three plain prompts — overhead, close-up, three-quarter view — and listed its visual choices as assumptions because the brief never describes the loaf.

Asked for the picture, it announced it was using the image skill and returned an overhead photo of a round, uncut loaf on a pale surface. Its note said the loaf's appearance was illustrative because no reference photo had been supplied. That is the right caveat: the picture is a guess at your product, not your product.

What Claude Opus 5.5 did

Opus 5.5 said plainly that it can't generate photographs in that chat. Instead it wrote an SVG illustration at 1080 × 1350 pixels — the right 4:5 shape — with a scored loaf, a slice and a blurred pumpkin, and it recommended a photo tool or, better, a photograph of the real loaf.

An illustrated SVG is useful as a layout mock-up. It is not a product photo, and Opus said so. Anthropic's announcement does not list image generation among Opus 5.5's outputs.

Making the picture in Ciyo

We pasted Opus's first prompt — the close-up with a cut slice and a pumpkin behind — into Ciyo's Image Generator with GPT Image 2.5 Flare at 4:5 and 1K. It cost 2 credits and returned an 896 × 1120 photo in well under a minute.

The image followed the prompt closely: scored crust, orange-tinted crumb, one slice, worn board, pumpkin top right, and no lettering. The planning model chose the details; the image model drew them.

• Earlier test: can GPT-6 generate images?

Which one to use for your post

Use Sol or Luna to plan: the brief, the list of shots, the caption and the list of guesses to check. If you are already in ChatGPT, asking Sol for the image is convenient, but you cannot choose the image model or its settings from there.

Use an image model you can steer for the final picture — ratio, resolution, quality and a reference photo of the real product. In Ciyo that is the Image Generator or the agent with GPT-6 Astra, which writes the prompt and places the image on the canvas in one turn.

What OpenAI said at launch

The launch post describes Sol and Luna as faster, cheaper models built on Astra's advances. It does not claim image output, which matches the model pages.