GPT Image 2.5 adds Sketch drawing interface and Flare, Sunburst APIs

• ChatGPT Images 2.5 adds a drawing-led Sketch surface and comment-based local edits, with minchoi's launch note showing the new drawing flow.
• The API now splits into fast Flare and edit-focused Sunburst, as pika_labs' API announcement puts both models live with transparent-background output.
• Early visual demos center on holding a scene together during changes: higgsfield_ai's lamp edit moves the light source and updates the mirror, while its stop-motion test holds a toy car's lighting and position from frame to frame.
• Reference images still need explicit constraints, as GlennHasABeard's comparison found that a lightly specified reference could drift while three locking lines held it closer to model.

Higgsfield's lamp test makes a tiny spatial edit carry through to the scene's reflected light. Its stop-motion clip holds a toy car's illumination over successive frames, and aakashgupta's prompt list turns that same constraint logic into a 12-part production prompt library.

OpenAI's launch announcement introduces Sketch as a way to draw directly in ChatGPT. Its developer announcement says the feature opens with @Sketch, alongside templates and comment-based edits that target a specific part of an image.

The API exposes two distinct 2.5 model IDs. OpenAI's image-prompting guide positions them this way:

• gpt-image-2.5-flare is the smaller, speed-optimized model, with image quality comparable to GPT Image 2.
• gpt-image-2.5-sunburst is the base, quality-optimized model, positioned above GPT Image 2 for demanding edits.
• Both generate and edit from text or image inputs, preserve subjects more precisely, and support transparent backgrounds.
• Sunburst offers low, medium, high, xhigh, max, and auto quality settings. Its model page lists $5 per million text-input tokens, $8 per million image-input tokens, and $30 per million image-output tokens.

Early tests break the claim into specific visual constraints:

• Exact wording, placement, and typography are the focus of higgsfield_ai's text test.
• A red sideboard's mirrored surface is the focus of higgsfield_ai's reflection test.
• A character cycles through settings and costumes in higgsfield_ai's character-consistency clip.
• higgsfield_ai's visual-continuity test calls the result its strongest continuity result so far.

Stable illumination across a toy-car stop-motion sequence

Cross-edit consistency outweighs speed for egeberkina, whose early reaction says faster generations are welcome but reliable edits are the draw.

Aakash Gupta's 12-prompt set treats the image as a set of named invariants, then requests only the permitted change:

1. Lock a character across future edits.
2. Keep a product fixed while changing the setting.
3. Edit one named region only.
4. Make ad variants by changing only the background.
5. Recompose to vertical without cropping the subject.
6. Place finished UI artwork inside a campaign scene.
7. Apply exact artwork to merch.
8. Swap an outfit while preserving person, pose, light, and background.
9. Replace furniture while preserving a room's structure and camera angle.
10. Localize a design while retaining layout and imagery.
11. Remove a flaw and rebuild the background behind it.
12. Turn a whiteboard sketch into a polished infographic while preserving boxes, arrows, and labels.

GlennHasABeard's Caturday reference test reports better reference handling than before. The stronger result depended on treating the reference as a fixed design, not just an inspiration image.

LinusEkenstam said early access helped create assets for Samurai Pippi Longstocking.

CharaspowerAI's character-sheet test starts from a single cyborg design and probes consistency across poses, angles, expressions, and costume details.

Ozan Sihay demonstrated a cross-tool motion workflow with GPT Image 2.5 and Grok Imagine Video:

1. Generate a café still as the opening frame.
2. Use GPT Image 2.5 to make a second still of the same woman walking onto the street, preserving the atmosphere.
3. Supply those images as start and end frames, then let Grok generate the movement between them.