ChatGPT Images 2.5 Guide: Multi-Turn Editing, Character Consistency and Flare vs Sunburst in the API

OpenAI's Images 2.5 cuts generation latency by up to 50% and keeps edits more consistent across turns. What changed from Images 2, how to use it for character-consistent series, how it compares with a Nano Banana pipeline, and what the API models cost.

PUBLISHED // UPDATED
DURATION // 06 MIN READ
Designer desk with a monitor showing a portrait being edited in small steps, a sketch stylus, and printed character sheets, no logos
FIG.01 // VISUAL DOCUMENTATION Designer desk with a monitor showing a portrait being edited in small steps, a sketch stylus, and printed character sheets, no logos

OpenAI released ChatGPT Images 2.5 on September 8, 2026. The update is less about a new visual style than about control: subjects from reference photos are preserved more faithfully, edits stay consistent over a longer conversation, and generation is up to 50% faster than Images 2.0. For illustrators, marketers and anyone producing a series of images around the same character or product, those editing improvements matter more than raw quality. This guide covers what changed, how to use it for consistent series, how it compares with a code-based Nano Banana workflow, and what developers should know about the two new API models.

What is new in Images 2.5

According to OpenAI’s announcement, Images 2.5:

  • produces more natural lighting and richer textures;
  • is better at preserving the subjects in reference photos;
  • edits “only what you’ve asked for” while keeping the rest of the image the same;
  • follows editing instructions more reliably across multiple turns, without degrading quality over time;
  • handles more complex layouts, including transparent backgrounds;
  • reduces generation latency by up to 50% compared with Images 2.0.

OpenAI says people create more than 3 billion images a week across ChatGPT Images and its API models. Images 2.5 is available on all ChatGPT tiers, including ChatGPT Work and Codex, on desktop, mobile and web.

For context, 9to5Mac notes that Images 2, released in April 2026, had already added 2K output, multiple aspect ratios, web research, better non-Latin text and two modes (Instant and Thinking). Images 2.5 builds on that base rather than replacing it.

New ChatGPT features

  • Sketch: draw a rough layout or outline directly in ChatGPT (type “@Sketch”) and use it as a guide for the final image.
  • Templates for common formats such as posters, flyers, merch and product photos.
  • Comments on images, so you can point to the exact area you want changed.
  • Prompt sharing, which lets others reuse a prompt with their own photos.

Why multi-turn consistency matters

The most common frustration with image generators has been drift: you ask to change one detail, and the model changes the face, the composition or the lighting too. After a few rounds, the image no longer looks like the one you started with. OpenAI’s claim for 2.5 is that earlier changes are “more likely to stay consistent” and each edit builds on the previous one.

In practice that makes ChatGPT more useful for:

  • Client revisions, such as “make the sky dusk” or “change the jacket to navy”, without starting a new composition.
  • Product imagery, where the product, background or copy changes but the brand treatment must stay.
  • Character series, where the same person or mascot appears in several scenes.

How to keep a character consistent in ChatGPT

A workflow that takes advantage of the 2.5 improvements:

  1. Create a character sheet first. Generate or upload a clear reference showing the character from the front, with distinctive features visible. Keep it in the same conversation.
  2. Describe changes as a short list of differences. For example: “Use the character from image 1. New scene: rainy street at night. Keep face, hair and coat identical.”
  3. Change one thing per turn when possible. The model handles focused edits best.
  4. Use comments to mark the exact region to change instead of describing its position.
  5. Start a fresh thread with the sheet if a long conversation begins to drift. Consistency is better, but not guaranteed indefinitely.

For a broader comparison of approaches, see our guide to consistent AI character design methods.

The code-based alternative: a Nano Banana pipeline

Developers who need repeatable results across many images may prefer an explicit pipeline to a chat thread. Google developer advocate Laurent Picard published a step-by-step notebook that takes an archive image, extracts a character into a new reference, then generates a sequence of scenes using Google’s Gemini 3.1 Flash-Lite Image (Nano Banana 2 Lite) model, the fastest and cheapest in that family.

The key design choice is that every generated image is stored as an asset with a record of which source images and prompt produced it. Each new scene explicitly passes in the character reference plus the relevant previous scene. The notebook also includes retry logic for rate limits.

ApproachBest forTrade-off
ChatGPT Images 2.5 in one conversationQuick revisions, small series, non-developersContext lives in the chat; harder to reproduce exactly
Scripted pipeline (Nano Banana or GPT-Image API)Large series, version history, automationRequires code, but every step is traceable and repeatable

API options: Flare and Sunburst

For developers, OpenAI released two models:

  • GPT-Image-2.5 Flare is the default for most applications. OpenAI says it delivers higher-quality images than GPT-Image-2 at 50% lower latency, suiting social content, product experiences, prototyping and high-volume generation.
  • GPT-Image-2.5 Sunburst targets premium workflows that need tighter control across edits, such as campaign creative, with longer generation times.

According to OpenAI’s Flare model page, Flare is billed by tokens at the same rates as GPT Image 2: $8 per million image input tokens and $30 per million image output tokens ($15 with Batch processing), plus $5 per million text input tokens. It supports the image generation and image edit endpoints, including inpainting.

Safety and provenance

OpenAI says Images 2.5 keeps its existing checks on prompts and images and continues to add C2PA metadata and invisible watermarking to identify images made with its tools. Teams that need to prove where an image came from should keep the original files with metadata intact; see our guide to AI art authentication and provenance tracking.

When to use something else

  • Document-style images such as forms, slides or dense text layouts may suit a model built for that purpose; see our look at Qwen-Image 3.
  • Files that must not leave your machine call for a local tool; we covered PhotoDirector’s local AI features.

Our analysis

Images 2.5 is an incremental release with a practical focus. The headline improvements, faster generation and edits that do not drift, address the issues that most often slow real creative work. Chat-based editing is now good enough for many revision cycles and small series, while teams producing large, reproducible sets will still benefit from a scripted pipeline that records every input. Test both on a short series of your own before choosing.