GPT Image 2.5: Flare vs Sunburst, Cost, Which to Use

§ 01What shipped#

01Every tier, every device: ChatGPT, ChatGPT Work and Codex, on desktop, mobile and web

02New in ChatGPT: Sketch, templates, comments on images, prompt sharing

03The API surface: two engines, five quality rungs, two backgrounds, exact size

1 / 3

←→

Two models, one direction. OpenAI’s split is by workflow, not by size. Flare “brings the same improvements in quality, editing, and speed to the API and is the default choice for most applications, delivering higher-quality images than GPT-Image-2 at 50% lower latency.” Sunburst “is built for premium visual workflows that benefit from tighter control across edits.” The launch post names Higgsfield AI, Adobe and Manus as early customers, with Higgsfield’s head of product praising Flare for “how well it understands what not to change.”

What OpenAI says improved. Sharper details, more natural lighting and richer textures; better preservation of the subjects in reference photos; edits that change only what you asked while keeping everything else, “even with more complex subjects and backgrounds,” and that stay consistent across multiple turns without degrading; more accurate real-world information in images; more complex layouts, including transparent backgrounds; closer adherence to a requested visual style. None of that is a benchmark. Arena had no 2.5 entry at launch, and OpenAI published no comparative scores.

§ 02The rumor scorecard#

As on our Astra tracker and Fable 5.1 tracker, we wrote the claims table below on September 4 and left it in place so it can be graded honestly.

The lesson matches Astra and Fable 5.1: the shipped build artifact was worth more than every anonymous benchmark sighting combined, and the price was never going to leak because it did not change.

§ 03What it costs#

The token rates are identical to GPT Image 2, so what an image costs depends only on the tokens each quality setting spends. We measured those tokens on both models at 1024 by 1024, five settings each; the numbers below are ours, not a vendor estimate.

The chart is in cents. Against GPT Image 2’s published tiers the ladder shifted a rung: 2.5 at high spends what GPT Image 2 spent at medium, and 2.5 at max spends what GPT Image 2 spent at high, so the two new settings are the top of the ladder, not a surcharge on the old rungs. Tokens do not scale with area either: a 3840 by 2160 image at low spends 371 tokens, under twice a 1024 square. The image generation guide carries OpenAI’s calculator; Batch API pricing for the two new models was not on the pricing page we pulled at launch, while the gpt-image-2 Batch rows still are.

§ 04Our read#

We route real image work every day: infographics, posters and product visuals produced inside larger jobs by CellCog’s image agent. Our watchlist from September 4 had two axes, text rendering and edit fidelity. Launch day answered the second loudly: precision editing is the headline of the whole release, and Sunburst exists for it.

We did not wait for a benchmark. On September 8 we ran Flare and Sunburst against each other: the same prompt (a product photo with one handwritten label), 1024 by 1024, all five quality settings each, through the Image API with usage reporting on.

The result: CellCog’s image agent gets the whole GPT Image 2.5 surface, with Flare at high quality as its default. Per request the agent picks Flare or Sunburst, one of five quality settings, a transparent or opaque background, and the size; the separate transparent-background tool that ran GPT Image 1.5 retires. Flare is the default because the two cost the same and Flare is faster; the agent reaches for Sunburst when a job is about edit precision, which is what OpenAI built it for. All of it lands under existing work with no change on the user’s side, and it ships as a product update. Every image on this page, the hero included, was made by that agent running Sunburst, from the same facts as the tables.

01The quality ladder as a race: same tokens, same price, Sunburst about 1.5 to 2x longer

02The cost ladder, measured: $0.006 at low to $0.211 at max per 1024 by 1024 image

03Five weeks, six signals: the rumor chain from mona-lisa-1 to launch day

04What OpenAI says changed in 2.5: six claims, no benchmark yet

1 / 4

←→

§ 05What we are still watching#

• Arena placement. gpt-image-2 at medium holds number one on text-to-image and both edit boards at launch; a 2.5 entry, and where Flare and Sunburst land relative to each other, is the first independent read.
• OpenAI’s own per-image table. The pricing page has token rates only; our measured ladder above stands until OpenAI publishes one.
• Batch API rows for the two new models.
• Text rendering. Our test prompt carried one handwritten label. A dense infographic pass against GPT Image 2 is next; results here.

§ 06Update log#

This is a living page; when the story moves, the update lands here.

September 9, 2026 - measured costs, more pictures. The inherited GPT Image 2 per-image table is gone; in its place our own measurement on both models at 1024 by 1024, 196 to 7,024 image output tokens, $0.006 to $0.211 per image, with 2.5 high costing what GPT Image 2 medium did. A three-picture strip on what shipped replaces a paragraph of tracker history.

September 8, 2026, evening - CellCog’s agent gets the whole 2.5 surface, Flare at high by default. Ran Flare and Sunburst at all five quality settings on one prompt: identical image tokens, Sunburst about 1.5 to 2x slower, no measured quality difference. CellCog’s image agent now chooses per request across both models, five quality settings, transparent or opaque backgrounds and size, with Flare at high as the default and Sunburst for edit precision; the separate GPT Image 1.5 transparent tool retires. New hero and a four-image slideshow, the first on this blog, all made by that agent running Sunburst.

September 8, 2026 - launched. OpenAI published “Introducing ChatGPT Images 2.5,” added gpt-image-2.5-flare and gpt-image-2.5-sunburst to the API changelog, model pages and pricing page, and started the ChatGPT rollout to all tiers. Page rewritten from rumor tracker to launch record; the September 4 claims table survives as the scorecard above. Verified against OpenAI’s launch post, both model pages, the changelog and the pricing page on September 8.

As of September 4, 2026: page opened. No official movement to log.