OpenAI has released two new image models under the name ChatGPT Images 2.5. Flare is designed for faster generation, Sunburst for more precise editing of existing images. The split thus follows two ways of working that actually look different in practice: sometimes many drafts in a short time, sometimes deliberately changing an existing motif.

However, it remains unclear which of the two models users get and when. OpenAI does not communicate this clearly, and initial tests suggest that not all accounts benefit equally from the improvements.

Two models, two work steps

For practical work in web projects, the distinction is worthwhile. The first step is usually exploratory: you need several variants of a motif to see whether an image idea holds up at all – for a header, a teaser tile or a category image, for example. Here throughput counts, not perfection. A model that delivers faster shortens precisely this loop.

The second step is different: a motif is basically in place, but the crop does not match the aspect ratio, a detail is distracting, an element should go. Such targeted interventions are traditionally the more difficult discipline with image models, because the model must keep the unchanged part of the image stable while recalculating only one area. A model optimised for editing addresses precisely this point.

Why model assignment is a practical problem

The fact that it remains unclear when which model is used sounds at first like a detail for enthusiasts. For teams that want to cast image production into a repeatable workflow, it is not.

  • Reproducibility: If the same prompt ends up with a different model depending on the plan or timing, results are only comparable to a limited extent between colleagues. Prompt templates that work for one person behave differently for another.
  • Expectation management: Anyone making a quality commitment in a project should know what it is based on. A demo created on a stronger model is not a suitable benchmark for day-to-day work in the team.
  • Plan dependency: The question of which quality arrives in which subscription also affects costing. It should be clarified before image production is planned as a standard step in an editorial process.

What this means for image production on the web

Generated images have long been more than a gimmick for websites – in many cases they replace the generic stock photo purchase for teasers, blog posts and landing pages. Two things remain important regardless of the model.

First: The bottleneck is rarely generation, but integration. An image must be available in several crops, compressed and delivered in modern formats, given a meaningful alternative text and sorted into the media management of the content management system. In TYPO3, for example, the clean maintenance of metadata and image variants determines whether an asset can be reused later. Faster generation only shifts the effort, it does not eliminate it.

Second: Editorial review remains. Image models continue to produce artefacts, especially with hands, text in the image, logos and technical details that professionals recognise immediately. For industries with products that need explanation, a generic but correct motif is often the better choice than a spectacular one with incorrect details.

A pragmatic approach

Anyone who wants to assess the new models cannot avoid their own tests – precisely because the assignment is unclear. A small, fixed test set makes sense: three to five typical motifs from your own project, each generated once anew and once specifically reworked, with the same prompts across different accounts. This shows faster than any announcement which quality actually arrives in your own setup.

In addition, it should be recorded which results go into live operation: prompt, model version as far as known, editing steps. This documentation costs little and is worth its weight in gold when there are questions about image rights or when switching models.

Assessment

The separation into a fast generation model and a precise editing model corresponds to how images are actually worked with in web projects, and accelerates the path from draft to usable visual. But as long as it is not transparent which model applies when, teams should formulate quality commitments cautiously and measure their own results instead of trusting announcements. The core of image work – cropping, formats, alternative texts, media maintenance in the CMS – remains unaffected by this in any case and continues to deserve the most attention.

Sources