GPT Image 2.5 Flare vs Sunburst: Pick the Still Desk, Then Image-to-Video

GPT Image 2.5 Flare vs Sunburst: Pick the Still Desk, Then Image-to-Video

ChatGPT Images 2.5 is OpenAI’s stills desk, launched 8 September 2026. In ChatGPT it is one consumer model. In the API it splits into GPT-Image-2.5 Flare (fast default) and GPT-Image-2.5 Sunburst (slower, tighter edits). Neither model outputs video. Lock the still, proof the type, then send only an approved frame to image to video or the AI video generator. ClipCanva is not affiliated with OpenAI, Runway, Adobe, or Canva. ClipCanva’s live stills page is still GPT Image 2. Do not treat a ChatGPT doodle as a finished pack shot.

Launch facts (8 September 2026)

Source	Public claim	Workflow implication
OpenAI: Introducing ChatGPT Images 2.5	Images 2.5 is the new ChatGPT stills model. OpenAI says it is sharper, better at keeping reference subjects, more precise on edits, and up to 50% lower latency than Images 2.0. Rolling out to ChatGPT, ChatGPT Work, and Codex on all tiers, desktop, mobile, and web.	ChatGPT users get one Images 2.5 desk. They do not pick Flare or Sunburst in the chat UI.
Same post	ChatGPT adds Sketch (@Sketch), templates such as Poster and Merch, on-image comments, and optional prompt sharing.	Sketch is a layout reference, not a legal still. Templates are starters. Shared prompts are remixes, not archives.
Same post	API ships two models: GPT-Image-2.5 Flare (quality + speed; default for most apps) and GPT-Image-2.5 Sunburst (extra precision, longer generation).	Name the API model in the job ticket. “Images 2.5” is not a model ID.
OpenAI: GPT-Image-2.5 Flare	Fastest everyday image model. Text + image in, image out. Quality: low, medium, high, xhigh, max, auto. IDs: gpt-image-2.5-flare and gpt-image-2.5-flare-2026-09-08. Token rates: text input $5 / cached $1.25 per 1M; image input $8 / cached $2; image output $30. Audio and video are not supported. Free API tier is not supported.	Draft volume, social stills, and first-pass product comps. Pin the dated snapshot if the pipeline must not drift.
OpenAI: GPT-Image-2.5 Sunburst	Most capable generate-and-edit model. Same quality knobs and the same token rates. IDs: gpt-image-2.5-sunburst and gpt-image-2.5-sunburst-2026-09-08. Speed is listed as medium.	Campaign keepers, label swaps, and multi-turn edits where the rest of the frame must not move.
Runway Dev changelog, 8 September 2026	gpt_image_2_5_flare and gpt_image_2_5_sunburst on Runway Dev. Prompt up to 32,000 characters. Up to 16 tagged references. 30 width:height ratios across 1K, 2K, 4K, plus auto. outputCount 1–10. Quality low / medium / high (default) / xhigh / max. Sunburst can return background: "transparent"; Flare is opaque or auto only. Billed 1–76 credits per image by quality and resolution, plus 1 credit per reference per generated image.	Runway’s meter is credits, not OpenAI tokens. Transparent cutouts are a Sunburst job on that API.

OpenAI also says Images 2.5 keeps C2PA metadata and invisible watermarking. Confirm live ChatGPT rollout, API org verification, and credit prices on the desk you actually pay. Token rates match GPT Image 2; OpenAI notes the GPT Image 2 calculator does not estimate 2.5 consumption.

Pick the desk before you generate

The fail is treating “make the product shot, then make it move” as one prompt.

Job	Where it actually runs	What you must supply	What the model may invent	Fail condition
Fast stills, many variants	ChatGPT Images 2.5, or API Flare	One-job brief, crop, lighting, exact on-frame text or “no text”	Wardrobe, background clutter, extra logos	Sending the first Flare to video because it looked sharp at thumbnail size
Precise edit of a keeper	API Sunburst	The approved still plus one change	Lighting drift, new type, extra fingers	Regenerating the whole scene because one SKU line was wrong
Layout doodle → concept	ChatGPT Sketch	Rough composition, then a written style lock	Product geometry, brand marks, readable claims	Using the doodle as the pack shot
Connected ClipCanva stills desk today	GPT Image 2	Prompt, optional references (live page: up to 3), aspect, resolution	Texture, type, small geometry	Claiming Flare or Sunburst is already the ClipCanva default
Design file + brand kit	Canva Magic Media / editor, not Images 2.5	Brand colors, empty type boxes, export size	Decorative type and fake logos	Shipping generated poster type as legal copy
Motion from an approved frame	Image to video	One still you already signed off, plus camera and duration	Camera path, extra limbs, new in-frame copy	Animating a draft that still has the wrong label
Spoken line not yet approved	AI script generator	Audience, one claim, forbidden words, close	Extra scenes, unearned superlatives	Prompting video before legal signs the line

Canva’s own split is a useful map, not a partnership. Canva Help says Magic Video sequences 1–10 clips or photos into a 60-second templated cut, Magic Media generates a short clip from a description inside a design, and Create a Video Clip is an 8-second cinematic clip with sound from a prompt. Those are finishing or generation desks. They are not Flare or Sunburst.

Flare vs Sunburst

Dimension	GPT-Image-2.5 Flare	GPT-Image-2.5 Sunburst
Official role	Fast default for most API apps	Premium generate-and-edit
OpenAI speed label	Very fast; up to 50% lower latency than GPT Image 2	Medium; longer generation
Best first job	Volume drafts, social stills, visual search, prototyping	Campaign creative, polished product imagery, multi-turn edits
Quality knobs	low … max, plus auto	Same list
Token rates (OpenAI docs)	Same $5 / $8 / $30 table	Same table
Runway Dev extras	Opaque or auto background	Transparent background supported
Do not use it for	Final label-legal pack shots you have not proofed	Cheap thumbnail batches

OpenAI quotes Manus that Flare ran two to four times faster than GPT Image 2 in their tests. Treat that as a customer measurement, not an OpenAI SLA. Higgsfield’s quoted win is “what not to change.” That is the Sunburst-shaped job even if you start on Flare.

Adobe Firefly and Runway are listed as early API customers. That does not make ClipCanva, Firefly, or Runway the same product. Credits, watermarks, and commercial terms stay with the desk that billed the still.

Draft-then-promote workflow

Write the packet once. Change only the variable. Promote only winners.

1. Name the output. Example: twelve 4:5 product stills, same SKU, English pack copy locked, then six 9:16 motion cuts.
2. Lock words off the image model. Claims, price, and CTA live in the AI script generator or a copy doc. On-frame text is a string you paste, not a vibe.
3. Draft on Flare or ChatGPT Images 2.5. One instruction per pass. Attach the product photo when identity matters. Use a prompt idea pattern if you need a reusable card.
4. Promote one still to Sunburst only when the change is surgical: swap the cap color, fix the headline, cut the background. Do not re-prompt the whole scene.
5. Human-gate the still. Spelling, logo, SKU, fingers, extra claims. Zoom to 100%. If ClipCanva is the stills desk you have today, run the same gate on GPT Image 2 or text to image.
6. Motion is a later desk. Approved frame → image to video. Multi-shot spots → AI video generator after scene cards. Compare paths on ClipCanva Compare when the renderer changes.

If the source is a webinar, invert the order: summarize on the AI video summarizer, verify quotes, then stills. Compression is a different job from invention.

Operator card for a 4:5 product still

Job: 4:5 ecommerce hero, English pack copy locked
Desk A: GPT-Image-2.5 Flare (or ChatGPT Images 2.5)
Input: approved pack-shot photo + one lighting brief
Prompt: same bottle, same label string, 50mm, window left, no extra logos
Quality: medium, then high only if type is already readable
Fail if: new claims, extra SKU, rewritten brand line

Desk B (only if Desk A is composition-right and one element is wrong):
GPT-Image-2.5 Sunburst
Change: [one instruction]
Keep: product geometry, label, crop
Fail if: full-scene regen, transparent background you did not ask for

Still: [filename]
Next desk: image-to-video, 4s push-in, no new type
Script: [filename] — VO is not in the still

On ClipCanva today, swap Desk A for the live GPT Image 2 controls. Do not invent a Flare toggle that the page does not show.

Creator checklist

• [ ] You named the stills desk: ChatGPT Images 2.5, API Flare, API Sunburst, or ClipCanva GPT Image 2.
• [ ] Flare vs Sunburst is written on the ticket if you are on the API.
• [ ] On-frame text is an exact string, or the brief says “no text.”
• [ ] Reference photos are files, not “the bottle from memory.”
• [ ] Sketch, if used, is labeled concept, not final art.
• [ ] You proofed spelling at 100% before any image-to-video job.
• [ ] Transparent cutouts, on Runway Dev, are a Sunburst request.
• [ ] Video is a second desk. Images 2.5 does not output video.
• [ ] Watermark / C2PA expectations match the channel you will publish to.
• [ ] You did not ask one model to invent the product, the legal claim, and the camera move.

FAQ

What is ChatGPT Images 2.5?

OpenAI’s 8 September 2026 product post names ChatGPT Images 2.5 as the new stills model in ChatGPT. It is positioned for sharper detail, more natural lighting, better reference-subject preservation, more reliable edits, and up to 50% lower latency than Images 2.0. It is rolling out to ChatGPT, ChatGPT Work, and Codex across all tiers.

What is the difference between Flare and Sunburst?

They are the two API models behind Images 2.5. Flare is the fast default for everyday generation. Sunburst is the slower, more precise generate-and-edit model. ChatGPT users do not pick those names in the consumer UI. Token list prices in OpenAI’s model docs are the same; you pay with time and retry count, not a different sticker price.

Does GPT Image 2.5 make video?

No. OpenAI’s Flare and Sunburst model pages list video as not supported. Generate the still, gate it, then use image to video or another video desk.

Is GPT Image 2.5 already on ClipCanva?

The live ClipCanva stills page is GPT Image 2. Treat the controls on that page as the source of truth. If you need Images 2.5 Flare or Sunburst today, use ChatGPT, the OpenAI API, or a host that lists those model IDs. Then bring approved frames back into ClipCanva for motion.

Should I skip Flare and always run Sunburst?

Not for volume. OpenAI calls Flare the default for most applications. Start there. Promote one keeper to Sunburst when the composition is right and a single element is wrong. Burning Sunburst on every thumbnail is how you wait longer for the same unproofed type.

Flare is the draft still. Sunburst is the surgical edit. Video is a later desk. Name the desk, lock the label, then generate.