Artificial Intelligence

Six image models shipped this month. Pick by job

November shipped a wave of image models, and the release trackers barely kept up. CSuite's November roundup lists six design-relevant releases in one month: OpenAI's GPT Image 2.6 Sunburst and Flare on the 4th, ByteDance's Seedream 5.5 on the 9th, MiniMax's H3 Image on the 10th, Microsoft's MAI-Image-2.7 on the 17th, Recraft's V4.1 Flash V2 on the 13th, and Topaz's Wonder 4 on the 3rd. Opper's release tracker feed confirms the GPT Image 2.6 dates independently. Six models in thirty days is a job market, and the practical question changes from "which model is best" to "which model does the job on your desk this week."

The job list

Transparent cutouts go to Seedream 5.5. CSuite's roundup flags its transparent-background output and steadier repeat edits, with API pricing from $0.02 to $0.035 per image. Anyone compositing product shots or building layered graphics knows the old workflow: generate, mask, refine the mask, repeat. A native transparent background deletes the mask step, and the mask step was the job.

Editable vector output goes to GPT Image 2.6 Sunburst and Flare. CSuite lists SVG export, animated GIFs of two to sixteen frames, and an adjustable complexity dial, priced at the previous generation's token rates. A generated poster that arrives as SVG lands in the same file format your brand assets already live in, which means the design team can open it and change things instead of starting over.

Character and product consistency across a set goes to MiniMax H3 Image. It accepts up to sixteen reference images and outputs at up to 20 megapixels, from about $0.02 per image. Sixteen reference slots is a moodboard with a render button attached.

High-volume drafts go to Recraft V4.1 Flash V2 at $0.007 per image, updated on the 13th. At that rate, eight candidate directions for a client meeting cost less than the coffee. Volume is a different job than final art, and pricing at half a cent per frame is what makes treating it as a separate job affordable.

Restoring photography goes to Topaz's Wonder 4, one-click restore and upscale up to six times, $0.0428 per image on Runware. Old campaign photography, archived brand assets, and scanned print work are the inputs here, not prompts.

The open-weights wrinkle

H3 Image ships open weights, 33.1 billion parameters and roughly 69 gigabytes, which sounds like the budget option for teams that self-host. CSuite's entry notes the license excludes the EU, the UK, Korea, and the US. Open weights under a region-excluding license are not open to most Western teams, and regional licensing just became a selection criterion alongside price and quality. Check the license before you plan the fine-tune.

What we kept from our own workflow

We generate nightly on Qwen Image 2.1 and swapped nothing this month, for reasons the wave clarifies rather than contradicts. Our rules are job rules: no generated type, because type lives in the layout layer; named palettes in every prompt, because palette is a brand parameter; fixed seeds with one varied slot, because consistency across a set is a prompt discipline. Swapping models changes the quality ceiling. Those rules survive every ceiling.

The swap triggers are feature-shaped, and two of this month's releases clear the bar. A native transparent background removes a compositing step we run weekly. SVG export removes the raster-to-vector conversion we run when a generated poster needs brand edits. Those two features map to real hours, so both earn a test in December. The rest of the wave is release-week noise until a job asks for it.

Pick the model by the job on your desk this week. The logos will rotate again next month.

Sources

The release dates, capabilities, and pricing in this post come from these trackers:

  • CSuite published the November roundup listing all six releases with capabilities, dates, and pricing.
  • Opper's release tracker feed confirmed the GPT Image 2.6 Flare and Sunburst dates of November 4 and the Gemini Omni 1.2 Flash release of November 5.

More in Artificial Intelligence