OpenAI's ChatGPT Images 2.5 Adds Sketch Input, Inline Comments, and Faster Generation
OpenAI announced ChatGPT Images 2.5 on September 8, 2026, introducing a new Sketch feature, inline image commenting, and up to 50% lower generation latency compared with Images 2.0. The release is available now for ChatGPT, ChatGPT Work, and Codex users on desktop, mobile, and web. The Verge
The headline addition is Sketch, which lets users draw a doodle directly inside the ChatGPT interface to guide image generation. Typing @Sketch into the chat box opens a drawing canvas where the user can make a quick sketch; that sketch is then used as a visual reference alongside the text prompt. This is a notable shift from purely text-driven image generation toward a hybrid input model, where rough spatial and compositional intent can be conveyed graphically rather than described in words.
ChatGPT Images 2.5 also adds the ability to leave comments directly on parts of a generated image to request specific changes. Rather than re-prompting from scratch or issuing a whole-image edit instruction, a user can point at a region and describe what should change there. For multi-turn editing workflows, this could reduce the iterative back-and-forth that has made precision image editing in chat-based tools cumbersome.
On the model side, OpenAI says Images 2.5 produces images with more natural lighting and richer textures. The company also says the new version is better at following editing instructions across multiple turns, addressing a known pain point where earlier image models would lose coherence or drift from the user's original intent as a conversation progressed. OpenAI's release notes independently describe ChatGPT Images 2.5 as bringing "sharper details, more precise editing, and faster image generation." OpenAI Help Center
The latency claim is the most quantifiable improvement: up to 50% reduction in image generation latency compared with Images 2.0. For interactive workflows where a user generates, reviews, edits, and regenerates, halving the wait time meaningfully changes the feel of the loop. It moves the tool closer to the responsiveness threshold where a user treats the output as a live draft rather than a batch-rendered result.
The broader context here is that OpenAI is iterating on image generation as a multidimensional problem, not just a quality race. The previous generation, announced in December 2025 under the title "The new ChatGPT Images is here," was described as powered by OpenAI's flagship image generation model, delivering more precise edits and consistent details. OpenAI Images 2.5 layers on top of that foundation with interaction-model changes (Sketch, inline commenting) and performance gains (latency, multi-turn instruction following) that are as much about workflow as they are about output fidelity.
The timing places this release amid a busy news cycle for OpenAI. The company's newsroom on September 8, 2026, also listed items spanning a Navier–Stokes solution proposed by an OpenAI model, quantum computing experiments run via GPT-5.6 Sol, and a safety post titled "An Alien Mind" published two days earlier. OpenAI Newsroom The breadth of that slate signals how OpenAI is positioning itself across research, applied AI, safety, and product, but Images 2.5 is the consumer-facing release in the batch.
For practitioners, the Sketch feature and inline commenting together suggest OpenAI is converging on an image-editing interaction model that resembles a lightweight creative tool more than a chatbot that happens to produce images. Sketch gives the model compositional grounding that text alone struggles to convey precisely — the exact position, relative size, or shape of an element. Inline commenting gives the user surgical control over specific regions across edit turns. Combined with the latency reduction, the practical effect is that iterative image workflows inside ChatGPT become faster and more directed.
In my view, the Sketch feature is the more interesting of the two interaction additions. Text-to-image models have been remarkably good at producing plausible single images from a prompt, but notoriously difficult to steer toward a specific composition. A sketch input is a direct, low-bandwidth way to communicate spatial intent, and it does not require the user to master prompt-engineering syntax to position elements. Whether Sketch produces reliable compositional control at scale, across diverse prompts and image styles, will be the practical test.
The multi-turn instruction-following improvement, if it holds up under use, addresses what has been the single most frustrating aspect of chat-based image editing: the tendency for a model to "fix" one thing while altering something the user wanted left alone. OpenAI claims Images 2.5 is better here. That claim will be tested quickly by the community.
