TechSifted has no affiliate relationship with OpenAI. This article is independent editorial coverage.
TL;DR: OpenAI released ChatGPT Images 2.5 on September 8, 2026. Generation latency is down up to 50% versus Images 2.0. Four new workflow tools landed: Sketch, Templates, Comments, and Prompt Sharing. Developers get two new API models: GPT-Image-2.5 Flare for speed and GPT-Image-2.5 Sunburst for precise control.
ChatGPT Images 2.0 launched in April 2026 as an architectural shift: it introduced a reasoning layer that let the model plan a composition before rendering it. That was the headline. Five months later, 2.5 is different in character.
The distinction matters for how you should think about the update. A reasoning model that runs with up to 50% lower latency while preserving quality is useful daily (Source). That is the core of what changed.
What Changed Under the Hood
The speed improvement is the most immediate fact. Per reporting from Unite.AI and tech-insider.org, generation latency is down by up to 50% compared to Images 2.0.
The quality improvements are subtler but important for anyone doing iterative work. Three things stand out from the reported changes:
Lighting and texture. The model now produces more natural lighting and richer textures. This matters most on photorealistic subjects, where Images 2.0 could produce outputs that looked accurate on structure but flat on surface. A portrait, a product shot, a material-heavy scene: those are the cases where this change will show up.
Reference photo fidelity. Images 2.5 better preserves the subjects in reference photos across multiple edits. This is a meaningful improvement for anyone using the model to maintain visual consistency across a set of outputs, which was consistently one of the harder things to get right.
Editing reliability. The model follows editing instructions more reliably across multiple turns of a conversation. That might sound like a minor behavioral note. For production workflows where a single image goes through six rounds of feedback, it is one of the more valuable changes in this release.
Images 2.5 is better at editing only what you’ve asked for, while keeping the rest of the details the same (Source). It also preserves background consistency and handles transparent elements more cleanly.
Four New Workflow Tools
OpenAI added four tools to the ChatGPT image interface with this release. They are aimed at different types of users and different stages of the creation process.
Sketch. Activated via the @Sketch command, this tool lets you draw a rough visual reference directly in the ChatGPT mobile app (Source). The model uses the sketch as a spatial reference alongside your text prompt. For anyone working on layouts, compositions, or anything with specific spatial requirements, this removes a step. You stop trying to describe a position in words and start drawing it. Creators and designers doing anything with composition-driven work will find this useful immediately.
Templates. Pre-designed formats for common output types: Poster, Merch, and product photos are the named examples in the release documentation. Each template includes guided follow-up questions to help shape the output. Marketers and social media teams producing high-volume assets will likely get more out of this than any other single addition. The friction in templated content production is usually not the generation itself but the back-and-forth to get format and layout right. Templates reduce that.
Comments. Direct annotation on generated images. You point to a specific area and describe what you want changed there. This is the most straightforward addition in the set. It creates a feedback loop that is more similar to how you would communicate with a human designer: point, describe, revise. Users doing brand work, UX mockups, or any task requiring precise spatial adjustments will use this constantly.
Prompt Sharing. Users can publish a prompt for others to recreate using their own reference photos and details. For teams, this enables standardized creative briefs that anyone can execute. For communities and creators, it opens a distribution vector: share a prompt that others can personalize and run with their own content.
None of these tools is transformational on its own. Together, they address a consistent gap in the 2.0 experience: the interface was powerful but friction-heavy for anyone doing production work. Templates and Comments in particular read as direct responses to enterprise feedback.
The API Split: Flare and Sunburst
For developers, Images 2.5 introduces two new models instead of one.
GPT-Image-2.5 Flare is the default. According to Unite.AI’s coverage, it delivers higher-quality images than GPT-Image-2 at 50% lower latency. This is the model you would use for high-volume generation pipelines, product catalogues, social media automation, or any workflow where throughput matters and you need quality without sacrificing speed.
GPT-Image-2.5 Sunburst is the precision option. Designed for production-ready workflows requiring exact editing control, it runs longer but delivers tighter results. Use cases where Sunburst is the better choice include anything requiring precise subject fidelity across reference images, detailed texture work, or high-stakes single outputs where you need the maximum available control.
Both models carry the same API pricing structure: $8.00 per 1M image input tokens ($2.00 for cached inputs), and $30.00 per 1M output tokens. Per-image costs at 1024x1024 resolution run from $0.006 at low quality to $0.013 at medium, $0.053 at high, and two new tiers: X-High at $0.094 and Max at $0.211. The X-High and Max tiers are new with this release.
The two-model approach is a practical split. Most product teams running image generation at scale will default to Flare. The smaller set of workflows requiring maximum output precision will justify Sunburst’s added generation time. Having both within the same API surface means you do not have to choose a single configuration for a mixed workload.
Where This Leaves the Competitive Field
The best AI image generators roundup we maintain reflects the current state of the category accurately: different tools are best at different things, and no single model clears the field.
Midjourney continues to lead on aesthetic output. For mood-driven creative work where the feel of an image is the primary objective, Midjourney produces images that are harder to match. That position does not change with Images 2.5. What 2.5 does is sharpen ChatGPT Images’ lead on the functional end of image generation: structured layouts, reference-based iteration, precise editing, production workflows with multiple revision cycles.
These are not the same market. A creator making artistic prints is working with different constraints than a marketing team producing fifty product images with brand-consistent backgrounds. The former may not move away from Midjourney. The latter has more reason to stay in ChatGPT Images after this update than before it.
Adobe Firefly occupies a distinct position that does not overlap directly. Firefly’s defining advantage is commercial IP indemnification. For brand teams in regulated industries or anyone producing work that will face IP scrutiny, that matters. The Firefly review covers this in detail. Images 2.5 does not change Firefly’s commercial safety position.
What Images 2.5 does is consolidate ChatGPT’s integration advantage. If your workflow already lives in ChatGPT, having a faster, more precise, more iterable image generation tool inside the same interface has compounding value. The new workflow tools reinforce this. You are not switching to a separate tool or managing a separate prompt library. Sketch, Comments, and Prompt Sharing are all native to the ChatGPT conversation model.
Who Should Pay Attention
Content creators and designers doing composition-driven work. Sketch addresses a real limitation in text-only prompting for spatial layouts. Regional editing and Comments make iterative feedback work the way it should.
Marketing and social media teams producing high volumes of branded content. Templates reduce setup time for recurring formats. The regional editing improvement means fewer full regenerations when a brief requires only a small change.
Developers building image generation into products. The Flare/Sunburst split gives you a meaningful performance/precision dial within a single API. The new pricing tiers for X-High and Max quality expand the output ceiling for production applications that need it.
For full context on ChatGPT’s non-image capabilities, the ChatGPT review covers what determines whether a Plus subscription is worth it at the tier level.
OpenAI reports users are now creating over 3 billion images weekly across ChatGPT and API platforms. That scale suggests the practical bar for image generation inside ChatGPT has cleared beyond early adopters. What 2.5 addresses is whether the tool holds up under the iterative, high-volume, production demands that come with that scale.
The answer is: more than 2.0 did.
ChatGPT Images 2.5 released September 8, 2026. Feature details based on reporting by Unite.AI and tech-insider.org. TechSifted has no affiliate relationship with OpenAI.