Skip to content
NewsProduct launch

OpenAI ships ChatGPT Images 2.5 with Sketch input, 50% faster generation, and two new API models

· by Pondero Newsdesk

The short version

OpenAI released ChatGPT Images 2.5 on September 8, cutting image generation latency by up to 50% versus Images 2.0 and shipping two new API model identifiers plus a freehand Sketch drawing tool for ChatGPT.

OpenAI ships ChatGPT Images 2.5 with Sketch input, 50% faster generation, and two new API models

On September 8, OpenAI published ChatGPT Images 2.5, cutting image generation latency by up to 50% versus Images 2.0 and introducing two new API model identifiers alongside a freehand drawing tool called Sketch. The release targets both ChatGPT users and developers building image pipelines on the API.

What

Four core improvements define Images 2.5 over its predecessor, per the OpenAI announcement: sharper detail and more natural lighting in generated images, better preservation of subject likeness when working from reference photos, more consistent style retention across multi-turn editing sessions, and more reliable handling of complex layouts including transparent backgrounds.

Latency is the headline number for production use. OpenAI says Images 2.5 generates results up to 50% faster than Images 2.0. Manus's evaluation team found the new Flare API model running "two to four times the speed of GPT-Image-2" in their internal tests, per the same announcement.

Four user-facing features shipped in ChatGPT at the same time.

Sketch lets users draw a rough layout directly in the ChatGPT interface by typing "@Sketch." The model reads that drawing as a compositional reference and incorporates it into the generated image. OpenAI positioned this as a way to communicate spatial arrangement or silhouette ideas without writing spatial descriptions in text.

Templates provide preset starting points for common output formats: flyers, product photos, and merchandise designs. Users select a template, then add specific content details.

Image comments let users annotate individual regions of an image before requesting an edit, so the model knows which element to change without relying on written location descriptions.

Prompt sharing lets users publish the exact prompt that produced an image alongside the image itself, so others can run the same generation substituting their own photos or details.

Images 2.5 is available to ChatGPT, ChatGPT Work, and Codex users across all subscription tiers on desktop, mobile, and web.

The two new API models

GPT-Image-2 shipped as a single model. Images 2.5 splits into two distinct API identifiers with different speed-quality profiles.

GPT-Image-2.5 Flare is the default. Per OpenAI, Flare delivers higher image quality than GPT-Image-2 at 50% lower latency. OpenAI positioned it for creator and social content, product imagery, visual search, rapid iteration, and high-volume generation workflows.

GPT-Image-2.5 Sunburst targets premium use cases where edit precision matters more than speed, per OpenAI's announcement. It carries longer generation times in exchange for tighter control across sequential edits. OpenAI cited production-ready campaign creative and polished product imagery as the intended use cases.

Adobe confirmed it is integrating both new models into Adobe Firefly. Matt Chotin, Senior Director of Product at Adobe, cited "faster generation and resolution consistency that keep images sharp and photorealistic through every refinement," per the OpenAI announcement. Runway's Chief Creative Officer Jamie Umpherson described Flare as fitting "seamlessly into how creators already work in Runway," also per that announcement. Higgsfield AI's Head of Product Axultan Alimkulov said Flare "understands what not to change" during edits, preserving "character, composition or visual identity" in a way that matters for film, UGC, and advertising workflows.

Pricing for the two new API models was not included in the launch documentation. OpenAI directed developers to its pricing page for rate details.

Why it matters

The Flare/Sunburst split is an operational change for any team routing image generation through the API. GPT-Image-2 gave developers one option, which forced teams to either accept its default latency or build workarounds through reduced quality settings. Two named models with explicit trade-off documentation lets engineering teams make that call cleanly at the call level. High-volume pipelines route to Flare; finalized asset production goes to Sunburst. That distinction also sets up cleaner cost attribution when teams analyze spending by workflow type.

The 50% latency reduction changes the math for real-time applications. At Images 2.0 speeds, in-session refinement across five rounds of editing was too slow for most interactive product contexts. At half the time per round, those same five edits become viable within a brief creative session. This crosses a usability threshold rather than nudging an existing one.

Sketch carries the most structural novelty. Text-to-image models have always struggled with spatial instructions: users writing "put the product in the upper left quadrant with the background fading right" typically get inconsistent results because language is a lossy medium for layout. A direct freehand input removes that translation layer. OpenAI has not announced Sketch API access. If that channel opens, developers could accept user sketches as generation inputs directly, removing the inpainting and masking infrastructure teams currently build themselves.

Context and reactions

OpenAI reported that users generate more than 3 billion images per week across ChatGPT Images and the GPT-Image API, per its announcement. That figure is a vendor self-claim and covers all tiers and channels combined.

The release follows GPT-6 Astra, which OpenAI published on September 3, and the company's August 31 announcement of a ChatGPT ad tier for free users. On the image generation API side, Google's Imagen 4 is the nearest direct competitor. Midjourney's v7 targets consumer and professional creative workflows. Neither company published a comparable latency reduction figure for their current models at the time of this article.

OpenAI says Images 2.5 includes C2PA metadata and invisible watermarking on generated outputs, consistent with its approach on prior image models. A system card for the release is available at deploymentsafety.openai.com.

What to watch next

OpenAI has not confirmed whether Sketch will reach the API. That announcement, if it comes, would change how developers design reference-image workflows in third-party products and remove a category of infrastructure workarounds.

Pricing for GPT-Image-2.5 Flare and Sunburst is absent from the launch documentation. Confirmation of separate rate cards, or parity with gpt-image-2, will determine whether the Sunburst premium is reflected in cost or purely in generation time.

Sources