I separate image tools into two workflows: exploring a visual idea interactively and generating assets from software. ChatGPT Images 2.5 covers the first; the developer-facing GPT-Image-2.5 models cover the second. Sharing a model family does not mean sharing billing or usage limits.
Standard ChatGPT Images 2.5 is available with a Free account on web, iOS, and Android, within the image-generation allowance shown in the product. The Flare and Sunburst APIs are usage-based products. A free ChatGPT account does not provide free API calls.
Start With the Billing Boundary
For occasional posters, product mockups, article graphics, presentation assets, and photo edits, I would start in ChatGPT. You can generate images in a conversation or open More > Images. There is no need to choose Flare or Sunburst from a model menu; those names primarily matter when integrating the API.
The Free-tier FAQ distinguishes image-generation limits from ordinary text-chat limits. OpenAI does not publish a permanent number of free generations that applies to every account. Product policy, demand, account conditions, and subsequent updates can affect the allowance. When you exhaust it, ChatGPT displays a notice; use the limit and reset information in your account as the authority.
I would not build a workflow around community claims such as “2–3 images per day” or “10 images every three hours.” Neither is a guaranteed Images 2.5 allowance. Also, standard image creation is separate from Images with thinking, which currently requires Plus, Pro, or Business.
The API has a different boundary. The Flare model documentation lists token-based pricing for text input, image input, and image output, and does not list Free-tier API support. Promotional credits may be available, but they are not a permanent entitlement.
What Changed From Images 2.0
OpenAI's September 8, 2026 announcement describes improvements to both image quality and the editing workflow: more natural lighting, richer textures, stronger preservation of reference subjects, and up to 50% lower image-generation latency than Images 2.0. That latency figure is a vendor-reported maximum improvement, not a promise for every request.
The improvement I care about most is preservation during edits. Changing a background, clothing detail, product element, or text should not require rebuilding the entire composition. Images 2.5 is designed for more precise local changes and better consistency across multiple editing turns.
The interface also adds workflow tools around generation: Sketch, Templates, comments, prompt sharing, reference images, selection-based editing, transparent backgrounds, and different aspect ratios. Compared with the more prompt-centric Images 2.0 workflow, the emphasis is on reusable layouts and iteration. Complex layout handling also improves, although that does not make the output deterministic design software.
Spend the Free Quota on Decisions, Not Ambiguity
My preferred prompt structure is asset + subject + composition + visual style + exact text + constraints + preservation rules. The point is to make the first request resemble a short creative brief, then refine the result. “Make a coffee poster” leaves almost every consequential decision unspecified.
For example: “Create a vertical promotional poster for a weekend coffee workshop. Use a warm cream and dark brown palette, soft natural shadows, a ceramic pour-over set in the center, and the headline ‘BREW BETTER COFFEE.’ Keep the lower 20% visually clean for event details.” That gives the model a subject, layout, palette, exact headline, and reserved area before spending a generation.
Supply Layout and Identity Up Front
Sketch is useful when placement matters more than descriptive prose. On supported experiences, type @, select Sketch, and draw the rough composition. Then give it rendering instructions: “Turn this sketch into a premium product advertisement. Preserve the placement shown in the sketch. Use realistic studio photography, a soft gray background, subtle shadows, and minimal typography.”
Templates handle another source of repeated prompting: canvas structure. ChatGPT includes starting points for posters, merch, and product photography. I would choose the format first, then specify the content: “Use this poster template for a minimalist AI developer conference. Headline: ‘BUILD WITH AI 2026.’ Add September 26, Tokyo. Use a black-and-white editorial layout, one abstract 3D object, generous negative space, and no extra copy.”
For recognizable objects or people, upload a reference instead of reconstructing their identity in text. Images 2.5 improves reference preservation for subjects such as products, portraits, branded objects, and interiors. A useful instruction is: “Keep the product shape, material, logo position, proportions, and camera angle unchanged. Replace only the background with a warm Scandinavian living room at sunset. Add realistic contact shadows and shallow depth of field.”
Refine the Existing Image
Once the composition works, I would stop requesting fresh versions. Open the image and use selection-based editing, or describe a focused change conversationally. “Change only the mug from white to dark green. Preserve the table, hands, lighting, camera position, typography, and every other object exactly as they are” makes the intended edit explicit. Preservation instructions help express intent; they do not guarantee pixel-identical surroundings.
The same approach applies to format changes: “Convert this into a 16:9 hero image. Keep the product at the same visual scale and extend the background naturally to the left and right. Do not change the product or typography.” Adapt the composition you already approved instead of describing it again from scratch.
For a more demanding initial brief: “Create a 4:5 studio advertisement for the uploaded sneaker. Keep the sneaker design and logo unchanged. Place it on a translucent acrylic platform with soft blue rim lighting and a dark charcoal background. Camera at product height, 70 mm commercial-photography look. Add the headline ‘MOVE DIFFERENT’ in clean white sans-serif type at the upper left. No other text or objects.” For isolated assets, explicitly request transparency, for example a centered silver satellite 3D icon with soft studio lighting and no text.
Pick the API Model by Workload
Both API models accept text and image inputs, produce images, support editing and reference-image workflows, and support transparent backgrounds. Both expose the quality settings low, medium, high, xhigh, max, and auto. ChatGPT manages these choices through its own interface rather than exposing the same developer controls.
gpt-image-2.5-flare is the speed-oriented choice for everyday generation, interactive applications, prototypes, and high-volume workloads. gpt-image-2.5-sunburst prioritizes precision-focused generation and editing, with longer generation time. I would start with Flare, then route requests to Sunburst when editing precision, reference fidelity, typography, or final asset quality justifies the tradeoff.
Move to an API when generation becomes part of the application: server-side requests, repeatable parameters, batch processing, automated edits, quality routing, usage tracking, or hundreds of calls. A paid ChatGPT plan addresses a different need: more interactive usage or paid-only functionality such as Images with thinking. Neither subscription status nor model naming removes API billing.
A Complete Generation Request
For a unified multi-model gateway, CometAPI offers both models through an OpenAI-compatible request pattern; any trial-credit offer should be checked against the current account terms. With the gateway key available in COMETAPI_KEY, the Flare request is:
curl "https://api.cometapi.com/v1/images/generations" \
-H "Authorization: Bearer $COMETAPI_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-image-2.5-flare",
"prompt": "Premium product photograph of a matte black wireless speaker on a light concrete pedestal, soft window light, realistic material texture, clean editorial composition, no text",
"size": "1536x1024",
"quality": "medium",
"output_format": "png"
}'
This requests a 1536x1024 PNG at medium quality. The credential, base URL, and model ID are the integration-specific pieces. I would keep this as the initial generation path and evaluate Sunburst for precision-critical work rather than defaulting every request to the slower model.
Read the Benchmarks as Early Evidence
The reported Arena results use human-preference battles, which provide a different signal from vendor benchmarks. Both GPT-Image-2.5 models are marked Preliminary, however, and have substantially fewer votes than the older GPT Image 2 baseline. The displayed results are:
| Arena category | GPT-Image-2.5 Sunburst | GPT-Image-2.5 Flare | GPT Image 2 |
|---|---|---|---|
| Single-image editing | 1520 ± 9 | 1491 ± 9 | 1461 ± 3 |
| Text rendering | 1481 ± 23 | 1434 ± 22 | 1427 ± 7 |
| Photorealistic / cinematic | 1423 ± 20 | 1406 ± 22 | 1381 ± 6 |
| Portraits | 1450 ± 32 | 1456 ± 35 | 1427 ± 8 |
Sunburst leads the displayed editing, text-rendering, and photorealistic results. Flare's portrait point estimate is slightly higher than Sunburst's, but the uncertainty matters. GPT Image 2 has more votes and narrower confidence intervals. I would use these results to select evaluation candidates, not treat small differences as permanent rankings or proof that one model wins every task.
Keep a Review Step Before Publishing
Images 2.5 still has limits beyond the free allowance. Complex briefs may require multiple iterations. Selection-based edits can affect pixels outside the highlighted region. Strict typography and pixel-level layouts can still need manual cleanup, so “preserve everything else” belongs in the prompt but should not replace inspection.
For commercial work, I would review important text, logos, product details, factual diagrams, and branded elements before publishing, especially at scale. OpenAI also says Images 2.5 retains C2PA metadata and invisible watermarking as provenance measures.
My decision rule is straightforward: stay with ChatGPT Free while manual creation fits the in-product allowance; consider a paid ChatGPT plan when those limits repeatedly interrupt work or you need Images with thinking; use the API when generation needs to run inside software. In every case, start with a well-constrained reference or layout, get the composition right, and spend subsequent requests on focused edits.
Originally published at cometapi.com
Top comments (0)