ChatGPT Images 2.5: faster generation, precise editing, and a new Sketch tool

ChatGPT Images 2.5: faster generation, precise editing, and a new Sketch tool

OpenAI has rolled out its newest image generation model to all ChatGPT users, promising 50% lower wait times and the ability to edit individual elements within images. Early independent testing against Google's Nano Banana 2 shows a mixed but mostly positive impression.

OpenAI launched ChatGPT Images 2.5 on September 8–9, 2026 — coverage differs slightly on the exact date (PCMag dates the launch to Tuesday, September 9, while an MSN-syndicated article gives September 8) — and the rollout is reportedly already underway for all users of ChatGPT, ChatGPT Work, and Codex on desktop, mobile, and web. The release comes just days after the company's GPT-6 Astra launch.

What OpenAI promises

All of the performance figures below are the company's own claims, relayed through coverage from outlets including PCMag and Craving Tech — no independent benchmarks of them exist yet:

• 50% lower wait times compared with the older model, according to OpenAI.
• Better rendering of natural light and rich textures.
• Stronger preservation of subjects from reference photos — meaning recognizable people or objects keep their facial and physical features consistent when placed into new settings, compositions, or styles.
• Better ability to follow editing instructions across multiple iterations.

New features in ChatGPT

The most concrete improvement for everyday users is precise editing: you can now swap a background without changing the subject, or edit text in an image while the rest is kept intact. The model also supports transparent backgrounds — a practical innovation for, among other things, sticker production, as Android Authority points out.

The new Sketch feature lets users draw directly in ChatGPT as a visual reference, activated by typing "@Sketch." In addition, ChatGPT is getting templates for common creative formats — such as "Poster" and "Merch" — where the user fills in information, design elements, and style. The company is also launching comments directly on images and shareable prompts.

Two models for developers

Through the API, developers get two variants: GPT-Image-2.5 Flare is positioned as the default choice for most applications, with OpenAI's quality promise: higher quality than GPT-Image-2 at 50% lower wait times (an OpenAI claim). GPT-Image-2.5 Sunburst is said to offer extra precision for detailed creative work, at the cost of longer generation times — but this description comes via a weaker source (an Instagram account quoting OpenAI's announcement) and should be read with caution.

The scale

According to OpenAI, more than three billion images are now created every week across ChatGPT Images and the GPT-Image models in the API. This, too, is a company claim without independent verification.

Early testing: against Nano Banana 2

Lifehacker's author ran a practical comparison against Google's Gemini Nano Banana 2. The conclusion is subjective but useful: the ChatGPT images seemed "a bit more real" on food — the author felt they actually looked more like real food than Google's — while Gemini did better on backgrounds and wider scenes. On precise editing (removing objects, changing time of day with rearranged cloud patterns), the ChatGPT result was "a bit more lived-in and a bit less artificial" than Gemini's.

PCMag's reviewers point out, meanwhile, an important caveat about Sketch: the prompt does "much of the heavy lifting" in the results — the doodle alone doesn't determine the outcome.

The road ahead

It's worth holding on to what is actually verified: no independent benchmarks have yet confirmed the 50% speed boost or the quality improvements, and OpenAI's own blog post is not available in the coverage underlying this article. Lifehacker's and PCMag's impressions are subjective, from single testers. But with immediate broad rollout — to all ChatGPT users, not just paying ones — this is a release that is easy to test yourself, which may be the most practical way to verify OpenAI's claims.