ChatGPT Images 2.5: Editing, Sketch, and API Models
ChatGPT Images 2.5 adds more faithful reference editing, stronger multi-turn consistency, Sketch, Templates, and Flare and Sunburst API models.
Before reading, try ChatGPT Images 2.5.
Developer: OpenAI · Released: September 8, 2026 · ChatGPT product: ChatGPT Images 2.5 · API models: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst
This version page covers the changes introduced with ChatGPT Images 2.5. For the model family's history and architecture, see the ChatGPT Image generation guide. For the preceding release's Reasoning Mode and baseline capabilities, see ChatGPT Image 2.
ChatGPT Images 2.5 is an incremental release focused on the part of image generation that happens after the first prompt: keeping a reference subject recognizable, changing only the requested element, and carrying earlier edits forward without gradual quality loss. OpenAI also added product tools that make the process feel more like a design workflow, including a drawable reference canvas, starting templates, comments on an image, and shareable prompts.
OpenAI says people create more than 3 billion images each week across ChatGPT Images and the GPT Image models in the API in its ChatGPT Images 2.5 release announcement. That figure is a statement about aggregate product usage, not a benchmark for model quality. Consumer-product features and rollout statements below refer to that announcement; API behavior is taken from the current OpenAI developer documentation.
What Changed From Images 2.0?
| Area | Images 2.0 | Images 2.5 | Why it matters |
|---|---|---|---|
| Reference fidelity | Reference-led generation and editing | Better preservation of recognizable subjects, lighting, and texture | Variations can stay anchored to a source photo |
| Editing precision | Conversational edits with the existing image context | Better isolation of the requested product, background, or copy change | A small revision is less likely to disturb the rest of the asset |
| Multi-turn consistency | Iterative editing in ChatGPT | Earlier edits are more likely to persist through longer conversations | Campaign and design revisions require fewer resets |
| Instruction handling | Complex prompts and visual reasoning | Stronger handling of complex information, layouts, transparent backgrounds, and style direction | Detailed briefs are more dependable |
| Generation latency | Images 2.0 baseline | Up to 50% lower latency, according to OpenAI | Faster iteration and higher-throughput workflows |
| Creative controls | Prompt and image input | Sketch, Templates, image comments, and prompt sharing in ChatGPT | More ways to express intent than text alone |
Source: the OpenAI release announcement cited above. The comparison describes the release delta; the latency statement is a release claim for ChatGPT Images, not a fixed API benchmark.
The important distinction is workflow reliability rather than a new architecture claim. OpenAI presents Images 2.5 as a model that is better at preserving what is already correct while applying a targeted change. That is especially relevant for product imagery, portraits, marketing variants, infographics, and other assets that go through several rounds of review.
Reference Fidelity and Precision Editing
Images 2.5 is designed to keep the identity and visual treatment of a reference while changing the setting, style, clothing, or composition. OpenAI specifically describes more recognizable subjects, more natural lighting and richer textures, and better retention of distinctive features. For an API workflow, this means a set of variants can remain tied to the same source image instead of drifting from one generation to the next.
Precision editing is the companion improvement. A user can ask to update one element, such as a product, background, or piece of copy, while preserving the subject, composition, and brand treatment around it. The useful test is not whether the model can make a new image; it is whether the requested change is local enough that the approved parts of an image remain approved.
This is an OpenAI capability claim, not a guarantee that every edit will be perfectly isolated. Small details should still be checked before publication, particularly faces, dense typography, logos, and images with overlapping objects.
Multi-Turn Editing Consistency
Long editing conversations create a different failure mode from one-shot generation: each revision can slowly alter earlier decisions. OpenAI says Images 2.5 follows specific editing instructions more reliably over multiple turns, keeps earlier changes consistent, and builds on the existing result without degrading image quality as quickly.
The practical workflow is straightforward:
- Generate a base image from text and references.
- Request a focused revision, such as a background, product color, or label change.
- Add another correction without re-uploading or rebuilding the whole brief.
- Review the final asset against the original reference and brand requirements.
The model's improvement reduces regeneration, but it does not remove the need for an approval checkpoint. Keep the original prompt, references, and final comparison in the production record when the image is used commercially.
Sketch, Templates, Comments, and Prompt Sharing
Sketch as a Visual Reference
Sketch lets users draw directly in ChatGPT and use that drawing as a guide for the final image. A rough room plan, an outfit contour, or a simple doodle can communicate spatial relationships that are difficult to describe in prose. The workflow is to draw, add a description of the desired style and details, and invoke Sketch with @Sketch.
Sketch is a product feature layered around the image model, not a separate API model. OpenAI's announcement demonstrates it for still-image creation. The API documentation covered on this page describes image generation and editing, not animated output.
Templates for Common Formats
Templates provide a starting point for common creative formats such as Poster and Merch. They move the user from a blank prompt to a format with an implied layout, after which the user can specify the information, design elements, and visual style. This is useful when the problem is not image generation itself but expressing a familiar deliverable clearly.
Comments and Shareable Prompts
Users can place comments directly on an image to identify the area that needs revision. When sharing an image, they can also share the prompt that produced it, allowing another person to reuse the idea with their own photos and details. These controls make review and remixing part of the same creation loop.
GPT-Image-2.5 Flare and Sunburst in the API
OpenAI's developer documentation lists two API models under the GPT Image 2.5 name. They should be treated as separate deployment choices, not as two ChatGPT subscription modes.
| API model | Official positioning | Best fit |
|---|---|---|
| GPT-Image-2.5 Flare | Smaller, speed-optimized model for fast, high-quality everyday generation; OpenAI's prompting guide describes image quality as comparable to GPT Image 2 | Creator and social content, rapid prototyping, and high-volume generation where latency is important |
| GPT-Image-2.5 Sunburst | Base model optimized for quality and OpenAI's most capable option for image generation and editing | Production-ready campaign creative and detailed edits where precision matters more than speed |
Source: OpenAI's GPT Image 2.5 prompting guide. The best-fit column is editorial guidance; measure quality and latency on your own workload.
The choice is therefore a workflow decision. Start with Flare when latency and throughput are binding constraints. Evaluate Sunburst when a slower generation is acceptable and the cost of a wrong edit or a drifting product detail is higher than the cost of waiting. Do not infer that Sunburst is universally better: OpenAI positions it for a narrower, higher-control use case.
For API integrations, use the exact OpenAI model IDs gpt-image-2.5-flare and gpt-image-2.5-sunburst. The VioEvo selector chatgpt-image-2-5 and its provider IDs are internal product identifiers, not OpenAI API model IDs.
Official API surface and output controls
| Area | Current OpenAI documentation |
|---|---|
| Image API | Use the model ID directly with image generation or image editing. Both Flare and Sunburst support text and image inputs and return image outputs. |
| Responses API | Use a supported mainline model at the top level and set Flare or Sunburst inside the image_generation tool. The image model is not a standalone top-level Responses model. |
| Quality | auto (default), low, medium, high, xhigh, or max |
| Size | auto or a custom WIDTHxHEIGHT; recommended sizes include 1024x1024, 1536x1024, and 1024x1536 |
| Background and format | auto, opaque, or transparent; output is png by default, with jpeg and webp also supported |
Custom dimensions must use multiples of 16, stay within 3,840 pixels per edge, keep the aspect ratio between 1:3 and 3:1, and contain between 655,360 and 8,294,400 pixels. Resolutions above 2,560x1,440 are experimental. For transparent assets, OpenAI recommends background="transparent" with PNG or WebP. JPEG and WebP support output_compression from 0 to 100.
Source: OpenAI's image generation guide.
Official API cost reference
Both GPT Image 2.5 models use the same token rates: text input is $5 per million tokens (cached input $1.25), image input is $8 per million tokens (cached input $2), and image output is $30 per million tokens. These are OpenAI API token rates, not VioEvo credits or a provider's per-image charge. Actual cost per image varies with prompt, input images, size, quality, and generated token usage.
Source: the OpenAI image generation guide cited above.
Availability and Safety
The release announcement says ChatGPT Images 2.5 is rolling out to ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web. The current developer documentation lists GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst for the API. Availability can still depend on the product surface, account, and organization.
For API requests, OpenAI filters prompts and generated images. The moderation parameter accepts auto (default) or low; blocked requests can return a moderation_blocked error with coarse input/output stage and category details. These checks do not replace human review of a deployed asset.
Known Limits and Evaluation Checklist
OpenAI's API guide lists several remaining limitations. Complex prompts may take up to two minutes; precise text placement, recurring-character or brand consistency, and layout-sensitive composition can still fail. Evaluate these areas before relying on 2.5 in production:
- Edit boundaries: Does a requested local change leave approved neighboring details untouched?
- Reference identity: Do faces, products, logos, and distinctive features remain recognizable across variants?
- Text and layout: Are all visible words, scripts, alignment, and transparent-background edges correct enough for the intended use?
- Long conversations: Does the asset remain consistent after several sequential edits?
- Latency trade-off: Does Flare's speed or Sunburst's tighter control better match the workflow's constraint?
- Safety handling: Does the workflow handle moderation blocks and review generated assets before publication?
This list is an evaluation recommendation, not an OpenAI benchmark. Proofread and inspect every final asset, especially commercial creative.
Frequently Asked Questions
What is new in ChatGPT Images 2.5?
The release focuses on more faithful reference images, more precise edits, stronger multi-turn consistency, richer handling of complex visual instructions, and a reported latency reduction versus Images 2.0. ChatGPT also adds Sketch, Templates, image comments, and prompt sharing.
What is Sketch in ChatGPT Images 2.5?
Sketch lets you draw a rough visual guide directly in ChatGPT and use it as a reference for the final image. OpenAI suggests room layouts, outfit contours, and doodles as examples; invoke it by typing @Sketch in ChatGPT.
What is the difference between Flare and Sunburst?
Flare is the speed-optimized choice for everyday generation, while Sunburst is the quality-optimized choice for demanding generation and editing. OpenAI recommends testing both against the same prompts, references, dimensions, and quality settings before selecting a production default.
Is ChatGPT Images 2.5 available in the API?
Yes. Use gpt-image-2.5-flare or gpt-image-2.5-sunburst with the Image API, or set one of them as the model inside the Responses API image_generation tool. Check current developer documentation for access requirements and pricing before integrating.
Does Images 2.5 guarantee perfect image edits or text?
No. OpenAI describes improvements, not a guarantee of perfect isolation or correctness. Review faces, logos, dense text, transparent edges, and every other production-critical detail before publishing.
Should I migrate an Images 2.0 workflow immediately?
Test the workflow first. OpenAI's recommended migration sequence is to save a representative baseline, test Sunburst when quality is the binding constraint, test Flare when speed is the priority, and keep prompts, references, dimensions, and quality settings controlled during comparison. Roll out by workflow only after quality, latency, failures, retries, and cost per accepted image are measured.
Use the image-to-image tool to test reference-led generation and iterative edits with ChatGPT Image 2.5 Flare.
