OpenAI Releases ChatGPT Images 2.5 With Sketch and Two New API Models

OpenAI released ChatGPT Images 2.5 on September 8, 2026, calling it the company’s new state-of-the-art image model, and added two companion models to its API. OpenAI said people create more than 3 billion images each week across ChatGPT Images and the GPT-Image models in the API.

Sharper Images and New Creation Tools

According to OpenAI, Images 2.5 produces more natural lighting and richer textures, better preserves subjects from users’ reference photos, and follows editing instructions more reliably across multiple turns of a conversation. The company said it reduced image generation latency by up to 50% compared with Images 2.0.

OpenAI said the model is better at changing only the elements a user asks it to edit while leaving the rest of the image intact, even with complex subjects and backgrounds, and that it renders images containing real-world information more accurately, handles more complex layouts including transparent backgrounds, and holds requested visual styles more consistently. During longer conversations, the company said, earlier edits are more likely to carry through and each new instruction builds on previous work without degrading image quality.

For API developers, OpenAI said the same fidelity makes reference-led workflows more reliable and allows a single element, such as a product, background, or line of copy, to be updated while preserving the surrounding subject, composition, and brand treatment.

Alongside the model, OpenAI introduced Sketch, a feature that lets users draw directly inside ChatGPT as a visual reference for the final image, invoked by typing “@Sketch” in a chat. New templates provide starting points for popular formats such as flyers and product photos, with named options including “Poster” and “Merch.” Users can also place comments directly on generated images for more focused editing, and can share the prompt behind an image so others can rerun the idea with their own photos and details.

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

Flare and Sunburst in the API

For developers, OpenAI released two image models in the API. GPT-Image-2.5 Flare brings the same quality, editing, and speed improvements and is positioned by the company as the default choice for most applications; OpenAI said it delivers higher-quality images than GPT-Image-2 at 50% lower latency, suiting uses from creator and social content to product experiences, visual search, rapid image prototyping, and high-volume generation. GPT-Image-2.5 Sunburst is built for premium visual workflows that benefit from tighter control across edits, such as production-ready campaign creative and polished product imagery, with longer generation times.

OpenAI published statements from three early customers alongside the release. Matt Chotin, senior director of product at Adobe, said OpenAI’s GPT-Image-2.5 models are now available in Adobe Firefly, Adobe’s creative AI studio. Lucky Liao of Manus’s evaluation team said Flare delivered high-quality images at two to four times the speed of GPT-Image-2 in the team’s evaluations, citing improved transparent-background generation. Axultan Alimkulov, head of product at Higgsfield AI, said the model preserves an original image’s character, composition, and visual identity through edits.

OpenAI’s pricing page lists both models at the same rates: image input at $8.00 per 1M tokens, or $2.00 per 1M tokens for cached input, image output at $30.00 per 1M tokens, and text input at $5.00 per 1M tokens, or $1.25 cached. The page describes Sunburst as built for workflows where editing precision matters most and Flare as OpenAI’s fastest model for high-quality, everyday image generation.

Safety Evaluations and Provenance

OpenAI also published a system card for the release on September 8, 2026. The card states that GPT-Image-2.5-Sunburst and GPT-Image-2.5-Flare were trained on diverse datasets spanning publicly available internet information, data accessed through third-party partnerships, and material provided or generated by users and human trainers, with filtering intended to reduce personal information and safety classifiers aimed at reducing harmful or sensitive content. It notes that heightened realism could, absent safeguards, enable more convincing deepfakes of real people, places, or events, which violate OpenAI’s usage policies.

The card describes a layered safety stack: LLM-based policy checks refuse violative requests before generation begins, and a safety reasoning model then moderates text and image inputs and inspects final outputs before they are shown, blocking generation when either violates policy.

In OpenAI’s automated adversarial evaluations, which used prompts designed to elicit policy-violating images across all tracked safety categories, final unsafe generations presented measured 1.09% for Sunburst and 1.41% for Flare, against a 1.64% baseline for ChatGPT Images 2.0. The card cautions that the adversarial test set is not representative of production traffic, that automated policy labels may contain errors, and that the findings apply to the fixed test set and configurations evaluated.

Under OpenAI’s Preparedness Framework, which tracks biological and chemical, cybersecurity, and AI self-improvement risks, the card reports that neither model crosses the Bio High or Cyber High thresholds. OpenAI said it nevertheless applies precautionary biological-risk mitigations, including an image-specific biological safety policy enforced through the safety reasoning model, live blocking at the input and output layers, offline review combining automated systems with human reviewers, and account suspensions where ongoing misuse patterns are detected.

On provenance, the card says OpenAI is continuing C2PA metadata through the C2PA Conformance Program and adding Google DeepMind’s SynthID invisible watermarking across ChatGPT, Codex, and the OpenAI API.