ChatGPT Images 2.5 combines faster image generation with targeted editing, while GPT-6 Astra is intended to handle more demanding work tasks. The new tools affect you if you create visuals for presentations, communications, or creative projects, or use ChatGPT as a work assistant. OpenAI introduced both offerings in September 2026, but questions remain about performance, cost, and availability.
What the new image tools offer
ChatGPT Images 2.5 is OpenAI’s new generation of tools for creating and editing images with Artificial Intelligence (AI). According to OpenAI, it should produce more natural lighting, richer textures, and sharper details. The system is also supposed to preserve subjects from reference photos more effectively and follow editing instructions more consistently over several conversational turns.
That matters most when you develop an image step by step. Instead of describing a new picture from scratch after every change, you can generate an initial scene and then adjust individual elements. In an example documented by Simon Willison, the model added a raccoon scientist studying a chart to an existing graphic. This kind of precise addition is more useful for presentations, training materials, or editorial illustrations than a system that rearranges the entire scene whenever you request a correction.
OpenAI offers two variants. Flare is positioned as the option for fast, high-quality everyday images. Sunburst takes longer but is intended to provide tighter control over edits and handle more demanding visual work. For regular ChatGPT users, however, it is not yet clear when each model is selected; The Decoder highlights this unresolved allocation.
OpenAI says more than three billion images are now created each week through ChatGPT Images and the GPT-Image models. This figure comes from the provider and has not been independently verified in the supplied sources. It nevertheless suggests why improvements in speed and editing could affect a large number of people.
Why Sketch lowers the barrier
Sketch lets you draw a rough picture directly from the prompt box and use it as part of your instruction. You no longer have to translate every visual idea into words. Lines, positions, and approximate proportions can be indicated visually, while a short text describes the desired subject and style.
In a brief test by The Verge, a cat doodled with a computer mouse was converted into a realistic image. The tester then commented directly on the relevant area to change the cat’s eyes from blue to green. This is a practical benefit if you find it easier to sketch spatial ideas than describe them, even when your drawing skills are unlikely to trouble a professional illustrator.
Step 1: Open Sketch
- Type @Sketch in the ChatGPT prompt box and open the drawing window that appears.
Step 2: Draw the idea
- Sketch the main shapes, then add a text instruction explaining which subject or visual style ChatGPT should develop from them.
Step 3: Correct individual details
- Comment on the relevant part of the image and describe the intended change as specifically as possible.
According to OpenAI, ChatGPT Images 2.5 and Sketch are available to ChatGPT, ChatGPT Work, and Codex users on desktop, mobile, and the web. The sources do not provide specific information about availability in Switzerland. They also do not clarify supported languages, the handling of uploaded reference images, or possible differences in data protection conditions.
What GPT-6 Astra promises for work
OpenAI describes GPT-6 Astra as its most capable model so far for business tasks. Generative Pre-trained Transformer (GPT) is the name of the model family that processes and generates text and other content. OpenAI emphasizes advanced reasoning, computer use, and stronger judgment for writing and design.
The available summary of the GPT-6 Astra announcement remains broad, however. It contains no benchmark results, prices, or details about which ChatGPT subscriptions include the model. The claimed improvements should therefore be treated as provider statements rather than an independent performance assessment.
One practical experiment at least illustrates how the image and work functions can be combined. Simon Willison first used ChatGPT Images 2.5 to create a photo of a Fabergé egg themed around a television show. He then passed the image to Codex running GPT-6 Astra and asked it to create a Blender file, meaning a three-dimensional model for the Blender 3D design application. The process took 17 minutes and 51 seconds and produced several files, which he inspected using a browser-based Blender viewer.
This provides a concrete professional example for design, product visualization, or early concept development: a two-dimensional draft can become the starting point for an editable 3D model. It was a single experiment under prepared conditions, not evidence that Astra can reliably turn any image into a finished 3D asset.
OpenAI places this development in a broader economic context. In a separate post, the company argues that more capable and affordable AI could expand the work people and businesses can complete while making growth more economical. This account of newly accessible work does not provide prices, case studies, or measurable productivity figures in the available summary.
Pros and Cons of the new OpenAI tools
Pros:
- Faster images – OpenAI claims up to 50 percent lower latency than Images 2.0 for the faster variant, meaning less waiting time before an image appears.
- More targeted corrections – Reference subjects, comments on image regions, and multiple editing turns may reduce the need to regenerate an entire image.
- Low barrier to entry – Sketch lets you communicate an image idea through a simple drawing instead of a lengthy description.
- Connected workflows – The Blender example suggests that Astra can help turn visual drafts into working files for later editing.
Cons:
- Provider claims – Key statements about quality, speed, and work performance have not yet been independently confirmed.
- Unclear model selection – OpenAI has not explained how ChatGPT users are assigned to Flare or Sunburst.
- Variable costs – Identical token rates, meaning prices for processed units of data, do not produce the same cost per image when token use differs.
- Unanswered privacy questions – The sources do not explain how reference images are handled in Switzerland or which terms apply to business files.
For the application programming interface (API), which allows other software to access the models, The Decoder reports rates of $8 per million image input tokens and $30 per million output tokens. A square 1024-by-1024 image costs about $0.006 at the “low” quality setting, approximately $0.053 at “high,” and around $0.21 at the new “max” level. These figures are not automatically the prices paid by people using ChatGPT directly.
Despite identical token rates, Sunburst can cost more per image than Flare, probably because it processes requests for longer. Images 2.5 also currently lacks a cheaper batch rate, which is a discount for processing many requests together. OpenAI does not provide an average cost per image, so a reliable budget depends on the actual workload.
What this means for you
If you are just getting started, Sketch is the most sensible first test. Begin with a simple subject, add a clear style description, and then change only one marked region. This quickly shows whether the tool preserves your original idea or alters unrelated parts during each edit.
For a workplace task, you could draw the rough layout of a presentation graphic and then refine its lighting, textures, or a single object. Only upload reference material that you are permitted to process with the service you are using. The supplied sources do not provide enough information to judge whether the tools are suitable for confidential company data or personal photographs in Switzerland.
As an advanced user, you may gain more by separating initial image generation from later refinement. You could create several quick variants with Flare and use Sunburst for precise editing, provided that model selection is available to you. For 3D or design projects, the Blender example offers one possible path from a reference image to a working file, but the resulting model, materials, and camera positions still require review.
GPT-6 Astra is most relevant when the intended result involves several work stages and files rather than a single answer. Without independent comparisons, the sources do not establish whether Astra completes such tasks more reliably or economically than other models. Specific prices and conditions for Swiss consumers, schools, and businesses are also absent.
ChatGPT Images 2.5 makes visual work more tangible: images should appear faster, sketches can replace part of a written description, and comments on specific regions simplify corrections. According to OpenAI, GPT-6 Astra extends this approach into more complex workflows, but the publicly documented examples remain limited. The largest unresolved risk is not one poor drawing, but uncertain costs, privacy conditions, and the reliability of the models during repeated professional use.
Sources
- Introducing ChatGPT Images 2.5 – Simon Willison’s Weblog, 2026-09-08
- ChatGPT Images 2.5: Schneller, präziser, aber nicht für alle gleich – The Decoder, 2026-09-09
- ChatGPT Sketch turns your bad drawings into detailed AI images – The Verge, 2026-09-08
- .blend URL Viewer – Simon Willison’s Weblog, 2026-09-09
- GPT-6 Astra: The next generation in intelligence for work – OpenAI, 2026-09-09
- The Work Now Within Reach – OpenAI, 2026-09-08
Image: George Milton via Pexels