OpenAI points Images 2.5 at the edits that used to degrade the rest of the frame
OpenAI leads on lighting and speed, but the claims that decide whether a team can ship an asset are about holding a subject consistent across turns and leaving untouched regions alone, and neither arrives with a number.
The Product Desk
What happened
- OpenAI is rolling ChatGPT Images 2.5 out to ChatGPT, ChatGPT Work and Codex users across all tiers on desktop, mobile and web, with both API variants open to developers.
- OpenAI says 2.5 holds detail the user did not ask it to change, including in busy scenes with several subjects, where Images 2.0 often degraded that detail during targeted edits.
- The company puts image generation latency at up to 50% lower than Images 2.0, the headline performance number in the release.
- A new Sketch tool lets users draw a rough layout by finger, stylus or mouse inside ChatGPT and have the model build a finished image on that reference.
- TechRadar reports that the Edit toolbar of Markup, Comment, Remove BG, Erase and Resize had already been arriving quietly in ChatGPT during August, with no formal announcement.
Why it matters
- Who has to decide decision Anyone building on the API now has to choose per asset rather than accept the default, because Flare is tuned for volume and Sunburst for tighter control across edits, and that routing rule has to be written down by someone.
- What it limits constraint Because every preservation claim is qualitative, a team cannot promise a client the same face twice on the strength of the announcement, and has to fund its own acceptance test first.
- What's now possible capability With each edit stored as its own version, a bad edit costs a scroll backwards instead of a restart from the first prompt, which is insurance that pays whether or not the stability claim holds up.
- Who's exposed exposure Assets leave with C2PA metadata and invisible watermarking attached, so a brand shipping this work is publishing a machine-readable provenance trail its clients can read.
Next story · Security · 5h ago
X attributes more than 13,000 suspended accounts to Iranian, Russian and Chinese influence operationsConfirmed2 publishers
TechRadar's reviewer describes the loop that came before this release: re-prompt an image that came back wrong, watch it come back wrong in a new way, then abandon it and start from the first prompt [13]. The two claims that speak to that loop are narrower than the release notes. OpenAI says identifiable details, including distinctive facial features, now carry across multiple generations and edits inside a single conversation [6], and that earlier edits stay stable as later ones are applied, each one building on the last [7].
Most of the rest is the pitch. More natural lighting and richer textures [3]. Templates for flyers and product photos, inline comments, and prompt sharing so a colleague can pick up your idea [8]. TechRadar, after 24 hours with it, says calling the feature surface overcrowded would be an understatement [15], and the genuinely useful thing it found sits outside the model: comments in the Edit toolbar work as instructions pinned to a point in the picture, carried out when you hit send [11]. That takes the job of saying which element you mean off the prompt and puts it on the canvas.
The speed numbers need a squint. OpenAI told 9to5Mac that Flare delivers higher-quality images than GPT-Image-2 at 50% lower latency [20], while SiliconANGLE reports the company describing Flare as between two and four times as fast as GPT-Image-2 [19]. Halving latency doubles the images you get per minute, 1/(1-0.5) = 2, which is the floor of that range and the only part of it the latency figure supports [21]. A capacity plan built on 4x will come up short.
What none of the reporting carries is a rate [17]. There is no published figure for how often a subject stops looking like itself by the fifth edit, and no benchmark for changes landing outside the region named in the prompt. That leaves the acceptance test to whoever owns the workflow, and it is a cheap one: ten reference photos of the subject you actually use, a five-edit chain on each, then two counts, subjects still recognisable at the end and changes that landed where nobody asked for them. The sort that matters is whether the subject has to stay identical across assets and whether an asset gets edited more than once. Only the corner where both are true needed this release. Single-pass social images were fine before it.
Who it is for is in OpenAI's own positioning: Sunburst trades speed for precision and is built for premium visual workflows with tighter control across edits, such as campaign creative and polished product imagery [10]. That describes the team producing the same mascot at forty sizes across a fortnight of revisions, which is also the team best placed to find out whether the edit-stability claim survives contact with a real brand guideline.
What to watch
- A published drift or collateral-edit rate, from OpenAI or an independent test, would turn the central claim from copy into a spec teams can hold a vendor to.
- Whether the Edit toolbar that TechRadar saw arriving in August is folded into the official 2.5 changelog, since the boundary between the two is currently guesswork.
- API pricing for Sunburst against Flare, which none of the launch coverage carries and which decides whether edit control is affordable per asset.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence45
- Adoption30
- Hype gap+30
- Incentives55
- Confidence60
Perspective Coverage
7 publishers- Builder
- Builder 22%
- Operator
- Operator 64%
- Investor
- Investor 14%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
A new tool called Sketch lets users draw a rough layout, visual direction or composition directly in ChatGPT with a finger, stylus or mouse, which the model then uses as a visual reference for generation.
- [2]
OpenAI announced ChatGPT Images 2.5, an update to ChatGPT Images 2.0, which debuted in April 2026.
- [3]
OpenAI says Images 2.5 produces more natural lighting and richer textures and follows editing instructions more reliably across multiple turns.
- [4]
OpenAI says it has reduced image generation latency by up to 50% compared with Images 2.0.
- [5]
With ChatGPT Images 2.0, details a user had not asked to change would often degrade when they attempted targeted edits; OpenAI says 2.5 better maintains the overall stability of the image, including in complex scenes with multiple subjects, objects or busy backgrounds.
- [6]
OpenAI says distinctive facial features and other identifying details can now be carried through more reliably across multiple image generations and edits within a single conversation.
- [7]
OpenAI says that over long conversations with multiple edits, earlier changes remain stable and each successive edit builds on them without loss of quality.
- [8]
Images 2.5 adds templates for common image formats such as flyers and product photos, inline comments on images for targeted edits, and the ability to share prompts alongside generated images.
- [9]
Images 2.5 is rolling out to ChatGPT, ChatGPT Work and Codex users across all tiers on desktop, mobile and web, with both Flare and Sunburst available to developers through OpenAI's API.
- [10]
GPT-Image-2.5 Sunburst sacrifices speed for greater precision and is built for premium visual workflows that benefit from tighter control across edits, such as production-ready campaign creative and polished product imagery.
- [11]
TechRadar found that comments in Edit mode are points for ChatGPT to act on: clicking an element and adding a comment such as 'remove this' or 'change this to blue' has ChatGPT follow the instruction when the user taps Send.
- [12]
TechRadar reports that in Edit mode every edit is saved as a separate file, so users can roll back to any previous version of an image by scrolling up or down.
- [13]
TechRadar describes the prior workflow as generating an image, re-prompting with a description such as 'make it brighter' or 'remove the people in the background', and after ChatGPT got it wrong a number of times, usually giving up and starting again.
- [14]
TechRadar says the new Edit toolbar, containing Markup, Comment, Remove BG, Erase and Resize, appeared to have been quietly rolling out in ChatGPT during August without OpenAI formally announcing it.
- [15]
TechRadar, which tested Images 2.5 for 24 hours, says there are now so many features in Images that to call it overcrowded would be an understatement, making the best new features hard to locate.
- [16]
OpenAI says Images 2.5 has guardrails on prompts and outputs and uses C2PA metadata plus invisible watermarking so generated images can be identified as such, to prevent unauthorised reuse and deepfakes.
- [17]
The reporting contains no measured rate for identity drift across edits or for changes to regions the user did not ask to alter; all such claims are qualitative statements from OpenAI plus one 24-hour hands-on account.
- [18]
The supplied launch coverage gives no API pricing for either GPT-Image-2.5 Flare or GPT-Image-2.5 Sunburst.
- [19]
SiliconANGLE reports that GPT-Image-2.5 Flare is the default model on opening the app and, per OpenAI, creates high-quality images at between two and four times as fast as GPT-Image-2, with better support for transparent backgrounds.
- [20]
OpenAI says GPT-Image-2.5 Flare is the default choice for most applications and delivers higher-quality images than GPT-Image-2 at 50% lower latency.
- [21]
A 50% latency reduction is a 2x speedup, so the two-times figure is the only part of the 'two to four times as fast' range that the latency claim supports.
Sources
7 independent publishers whose own reporting we read for this story.
- 9to5mac.comOpenAI releases ChatGPT Images 2.5 with ‘sharper details’ and ‘more precise editing’
2 articles · September 8, 2026
- androidauthority.comChatGPT can now create AI slop posters with more accuracy - Android Authority
1 article · September 9, 2026
- digitaltrends.comChatGPT Images 2.5 lets you sketch what you want instead of trying to describe it - Digital Trends
1 article · September 9, 2026
- siliconangle.comOpenAI’s ChatGPT Images gets faster, with sharper details and more refined edits
2 articles · September 8, 2026
- techradar.comChatGPT Images 2.5 is out — I’ve been testing it for 24 hours, and these are the 3 new features you’ll actually use | TechRadar
1 article · September 8, 2026
- theverge.comChatGPT Sketch turns your bad drawings into detailed AI images
1 article · September 8, 2026
- xda-developers.comChatGPT Sketch turns your paint-like doodles into beautifully rendered images
1 article · September 8, 2026