Breaking

OpenAI Releases GPT-Image-2.5 Models With More Precise Editing

September 8, 2026 at 22:39 EDT

Foundation Models
Media Generation

OpenAI has released GPT-Image-2.5 as ChatGPT Images 2.5, with a new focus on precise, localized image edits. A Higgsfield evaluation highlights the practical target: moving a lamp to the right while leaving the rest of the scene intact, a kind of change earlier models struggled to execute reliably.

September 8, 2026 · OpenAI

GPT-Image-2.5 makes precise edits the headline feature

OpenAI's new image models focus on changing one element while preserving the surrounding scene, with faster iteration and stronger multi-turn consistency.

The release at a glance

Up to 50%lower latency than Images 2.0

2–4×Flare speed in a Manus evaluation

2new API model variants

$30per 1M image-output tokens

Latency reduction, shown to scale

At the maximum claimed reduction, Flare's remaining latency is half the Images 2.0 baseline.

100%

Images 2.0 baseline

50%

Flare at max reduction

What improves — and what remains

Control gains

Targeted edits preserve surrounding detail
Multi-turn changes hold identity and composition more reliably
Transparent-background generation is improved

Trade-offs

Sunburst takes longer for tighter control
The API is image-only, without audio or video support
More realism increases the deepfake safety burden

The editing threshold

The differentiator is what stays unchanged.

GPT-Image-2.5 is aimed at the repeated revisions that turn a generated image into a finished asset.

Change one elementPreserve the sceneIterate with control

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.