GPT Image 2.5: Flare vs Sunburst, Sketch Input and Prompt Patterns
Short answer: GPT Image 2.5 (September 8, 2026) is GPT Image 2 with a faster lane and better tools. Use Flare to explore and Sunburst to finish. Draw the layout with @Sketch instead of describing positions, pin small edits with an in-image comment, and write the prompt as a director's brief — no negative field, no tag soups.
What is new
- Two API models. Flare (fast, the default) and Sunburst (precision).
- Sketch input. Draw a composition in the chat and reference it as @Sketch — the model uses it for layout, your text for content.
- Templates and shared prompts. Reusable briefs for a series, shareable with a team.
- In-image comments. Click a region, write the change, keep the rest.
- Reference preservation. A real face or product from a photo survives edits more reliably than in GPT Image 2.
- Transparent backgrounds and up to 50% lower latency.
Which model for which job
| Task | Model | Why |
|---|---|---|
| Drafts, thumbnails, dozens of variations | Flare | Fastest, cheapest per iteration |
| Final hero image, packaging, poster | Sunburst | Precision, text fidelity, detail |
| Keep a real person's likeness from a photo | Sunburst | Best reference preservation |
| Layout from a rough drawing | Flare + @Sketch, then Sunburst | Sketch fixes composition; Sunburst finalises |
| Transparent-background cut-outs | Either | Native transparent output |
| Small focused edit ("only the sofa") | Flare with an in-image comment | Comment pins the region |
Prompt patterns that work
- Generation: "A single photorealistic product photograph of a matte black ceramic mug on a light oak table, soft north-window light from the left, 50 mm lens at eye level, shallow depth of field, no text, no reflections of a person."
- Edit with a reference photo: "Use the attached photo for the person's face, hair and body only — do not copy its pose, background or lighting. Place her in a sunlit café, three-quarter view, natural skin texture, keep the face exactly."
- Sketch-driven layout: "Follow @Sketch for the placement: headline top-left, product centre-right, a small badge bottom-left. Render as a clean e-commerce banner, 16:9, brand colours navy and cream."
- Focused edit: comment on the sofa and write "replace with a linen sectional in warm grey, same size and position, keep everything else".
- Text in the image: keep copy short, put it in quotes, verify at full size — GPT Image 2.5 is one of the best at typography, but proofread anyway.
What to drop from your old prompts
Negative-prompt fields (write constraints instead), quality tag soups ("8k, masterpiece, trending on artstation"), and positional paragraphs that @Sketch now replaces. Keep one subject, one light, one mood, and one clear instruction per edit.
Pricing and labeling
The API bills tokens: about $8 per million image input tokens, $30 per million image output tokens and $5 per million text tokens; community estimates land a 1024×1024 image at roughly $0.006 (low), $0.053 (medium) or $0.211 (high). OpenAI attaches C2PA content credentials to outputs — keep them, because the EU AI Act's transparency rules (Article 50) have applied since August 2, 2026. For the wider field see the best AI image generators in 2026.
FAQ
What is GPT Image 2.5?
OpenAI's image model released on September 8, 2026 as "ChatGPT Images 2.5". It keeps the instruction following and editing strength of GPT Image 2 (April 2026) and adds Sketch input, reusable templates, shared prompts, in-image comments for focused edits, better reference-photo preservation, transparent backgrounds and up to 50% lower latency.
Flare or Sunburst?
Flare is the fast default API model — drafts, thumbnails, volume, quick edits. Sunburst is the precision model — final assets, text-heavy layouts, product shots and anything where a reference face or a brand element must survive exactly. A common pattern is Flare for exploration, Sunburst for the final render.
What is Sketch input?
You draw a rough composition directly in the chat (or send a sketch image) and reference it with @Sketch. The model treats it as a layout reference — where things go — while the text prompt defines what they look like. It replaces long positional descriptions.
Does GPT Image 2.5 support negative prompts?
No separate negative field. Write exclusions as plain constraints ("clean white background, no text, no reflections"). Like the other 2026 frontier models it reads a director's brief, not a tag list.
How is it priced?
Token-based on the API: roughly $8 per million image input tokens, $30 per million image output tokens and $5 per million text input tokens. Community estimates put a 1024×1024 image at about $0.006 (low), $0.053 (medium) and $0.211 (high quality); check OpenAI's pricing page for the exact tier before budgeting a batch.
How does it compare with Nano Banana Pro and Seedream 5.0 Pro?
GPT Image 2.5 leads instruction following and text rendering; Nano Banana Pro leads reference-driven generation (up to 14 reference objects) and conversational editing; Seedream 5.0 Pro leads pixel-level interactive editing. Many teams keep all three and route by task.
Want GPT Image 2.5 briefs written from a few clicks? GoldenPrompts builds literal, single-instruction photo prompts tuned for GPT Image, Nano Banana Pro and Seedream 5.0 — reference roles included. Free to start: 24 hours of everything, no card.