FULL STORY

GPT-Image-2.5: From Leak to Launch and Hands-On Tests

Days after leaks, OpenAI quietly launched ChatGPT Images 2.5 on Sept 9; its Flare and Sunburst variants swept image leaderboards, with users praising faster generation and striking character consistency.

2026-09-07 ~ 2026-09-10 · 6 episodes · 81 posts

Episode 1 · GPT-Image-2.5 Reportedly Launches in ChatGPT, Unconfirmed by OpenAI (2026-09-07, 5 posts)

From September 7–9, multiple X users and a developer reported on the rollout of OpenAI's next-generation image model GPT-Image-2.5, with some claiming it had already launched — though none of the reports have been officially confirmed and the accounts conflict.

Confirmed

  • Developer andpoul stated that gpt-image-2.5 is not yet visible to him in the API, and that it is hard to tell which model ChatGPT actually calls; he asked whether anyone had gained access
  • @realsohamparekh posted that the model is "coming soon," without details or sources

Unconfirmed

  • On September 7, @markk claimed the model could launch in ChatGPT on Thursday, was in release preparation, and was being quietly tested with some users; expected improvements include faster generation, better quality, fewer artifacts and noise
  • On September 9, @Angaisb and @eyishazyer claimed the model had launched: sharper detail, more natural lighting and textures, better identity consistency with reference photos, more precise editing, no drift across multi-turn edits, latency reduced by up to 50% versus Images 2.0, and two API variants; @eyishazyer added it was available to all paid tiers and free users
  • All launch claims come from ordinary accounts with conflicting timelines ("Thursday" vs. "already live"), so version, latency figures and availability remain unverified

Why it matters

If true, GPT-Image-2.5 would be a major upgrade to ChatGPT's image capabilities — halved latency and stable multi-turn editing directly improve UX, while dual API variants could affect developer cost and selection. Until OpenAI confirms, treat these reports skeptically.

Episode 2 · OpenAI launches ChatGPT Images 2.5 with Flare and Sunburst API models (2026-09-09, 43 posts)

On Sept 9 OpenAI released ChatGPT Images 2.5 to all ChatGPT, Work and Codex users across desktop, mobile and web, plus GPT-Image-2.5 Flare and Sunburst in the API. The release emphasizes speed, fidelity and controllable editing, adding Sketch input and templates — though, as @kimmonismus noted, it was overshadowed by OpenAI's same-day Navier–Stokes breakthrough announcement.

Confirmed

  • Generation latency down up to 50% vs Images 2.0; ManusAI (via @OpenAIDevs) measured 2-4x faster output than GPT-Image-2
  • Higher fidelity: more natural images, better person recognizability, richer textures, complex layouts and transparent backgrounds
  • Better consistency of people, pets and details across multi-turn edits, with per-edit version history
  • Comment-based local editing: change only what you want via point/circle selection; an editing toolbar (annotations, comments, background removal, eraser, zoom) was also reported via @rohanpaulai
  • New Sketch feature: type "@ Sketch" to draw a compositional reference or doodle onto images; users also found a canvas feature turning sketches into polished images
  • New templates for posters, merch and other common formats
  • API: Flare (default) is 50%+ faster with high quality; Sunburst targets fine editing with sharper detail, better lighting and stronger control
  • LMArena feedback (via @DavidmComfort) notes a big leap in in-image text; a user demo (via @ZeroStateReflex) showed it passing the scrambled Rubik's cube mirror test; per @calabiandyau, Sunburst topped three LMArena leaderboards with 72h free Arena access
  • Greg Brockman (gdb) shared the announcement and an image 2 vs 2.5 comparison, praising the capability

Why it matters

  • Comment- and selection-based editing lowers the barrier to precise retouching, pushing image generation toward iterative workflows
  • Templates, Sketch and canvas features lower the entry barrier for non-professional users, signaling a more mass-market creative tool direction
  • Being overshadowed by the Navier–Stokes news on launch day reflects the sheer density of OpenAI's releases

23 more related posts →

Episode 3 · ChatGPT quietly upgrades to GPT-Image-2.5 with new Sketch feature (2026-09-09, 20 posts)

On September 9, users discovered that ChatGPT's image generation model had quietly been updated to GPT-Image-2.5, with no major official announcement from OpenAI, alongside a new Sketch feature. Community testing concludes the update is a modest but focused improvement: text rendering, local editing, and sketch-to-image are clearly better, while long-standing texture noise issues remain largely unchanged.

Confirmed

  • GPT-Image-2.5 is live in ChatGPT, confirmed by multiple users including markk and xiaohu, plus a user test shared via DeryaTR.
  • OpenAI shipped Sketch alongside Images 2.5: typing @Sketch in ChatGPT lets users hand-draw a sketch that the AI turns into a finished image; currently web-only, not on desktop. APPSO tested it with a wobbly movie-poster sketch; aitrendzxyz shared a test turning a few irregular lines into polished images across four creative directions; Andrew Mayne (shared by lukaszkaiser) argues such small design-adjustment tools deliver outsized productivity gains.
  • markk ran the classic "noise artifact" test (a lush, high-detail forest clearing photo) with mixed results; he concluded noise is no worse but no better—the new version is essentially flat on texture artifacts.
  • Text rendering is a clear win: a Japanese designer found text in generated webpage designs no longer garbled, and Ben Geskin produced an "iPhone Duo - Apple Event" slide deck with crisp text.
  • xiaohu called the new model "something" and demonstrated precise local edits on infographics without breaking overall layout, plus layout adjustment, with side-by-side comparisons against other models.
  • Other tests: ZeroZJQ showed the new model handles the "slumped against a wall" pose that many models fail; GodsBurnt generated a product shot with lighting, equipment, and even the barcode on a graded card; Mayz1169 shared prompts converting any character reference into a 4×4 pixel battle sprite sheet (tested with Tanjiro) and pixel idle animations. aitrendzxyz summarized the three core improvements as sketch-to-image, targeted local editing, and readable text.
  • Aizkmusic tested with a character-selection scene, calling it a slight, "half-generation" upgrade whose outputs still visibly look AI-generated—"this is its worst moment," and it will only get better.

Unconfirmed

  • ElonBaldMusk complained the patterned noise is "worse than before," conflicting with markk's "no worse but no better" verdict; the artifact verdict remains unsettled.

Why it matters

  • Image generation is a core ChatGPT capability, and the "2.5" naming signals incremental iteration rather than a major leap. Community views on artifact fixes are cautious, but Sketch interaction, precise local editing, and reliable text rendering show markedly better controllability over structured graphic content—valuable for UI mockups, slides, and product shots—and worth tracking as official release notes and more comparisons emerge.

Episode 4 · GPT-Image-2.5 Variants Sweep Major Image Arena Leaderboards (2026-09-09, 6 posts)

After OpenAI released ChatGPT Images 2.5, its two variants GPT-Image-2.5-Sunburst and Flare quickly swept the major image arenas, taking the top two spots on all three LMArena image leaderboards and Design Arena's Image Editing Arena, with Sunburst subsequently made available in Arena Direct Mode.

Confirmed

  • LMArena officially announced that Sunburst and Flare rank first and second respectively across the text-to-image, single-image editing, and multi-image editing arenas, both significantly surpassing GPT-Image-2 (medium), and also sweeping the top two in all subcategories such as 3D modeling and art
  • A repost noted Sunburst's multi-image editing score surged by 81 points
  • Design Arena announced that GPT-Image-2.5 took the top two spots on the Image Editing Arena: Sunburst leads with Elo 1386, Flare second at 1360, with editing speed 6.2x faster than before
  • LMArena made GPT-Image-2.5 (Sunburst) available in Direct Mode, letting users interact with the model directly without blind battles; Leaderboard and Max modes are also offered

Why it matters

  • Two new variants topping multiple major image leaderboards signals OpenAI's clear edge over competitors in image generation and editing
  • The 6.2x editing speed gain has direct value for production workflows
  • Direct Mode availability lets users experience the strongest version immediately, accelerating adoption

Episode 5 · OpenAI Ships GPT-Image-2.5 Flare and Sunburst via API (2026-09-09, 4 posts)

OpenAI released two GPT-Image-2.5 models via API: the faster Flare and the higher-quality Sunburst, with Sunburst currently API-only and pricing largely unchanged from GPT-Image-2.

Episode 6 · GPT Image 2.5 Impresses Users With Character Consistency (2026-09-09, 3 posts)

Early hands-on reports say GPT Image 2.5 is faster and better at prompt adherence than 2.0, with striking character and background consistency across scenes—though some examples turned out to be elaborate Rickrolls.