OpenAI Launches GPT-Image-2.5 in Two Variants, Cuts Latency Up to 50%

OpenAI has released two new image models, GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst, replacing GPT-Image-2 as the default image generation and editing systems in ChatGPT and the API. OpenAI claims the faster of the two, Flare, cuts image generation latency by up to 50 percent compared to its predecessor while improving image quality.

Two models, one price

Flare is positioned as the default option for most use cases, offering higher quality than GPT-Image-2 at roughly half the latency. Sunburst is built for more demanding editing work, trading generation speed for tighter control over edits, according to OpenAI.

Both models share identical API pricing: $8 per 1 million input tokens and $30 per 1 million output tokens. But actual cost per image varies because token consumption differs by model and by the new quality tiers OpenAI introduced alongside the release — "xhigh" and "max" — which sit above the previous ceiling of "high."

For a 1024x1024 image, the "low" tier costs about $0.006, unchanged from GPT-Image-2. The "high" tier runs approximately $0.053. The new "max" tier costs roughly $0.21, consuming around 7,024 output tokens — the same price point as GPT-Image-2's old "high" tier. Unlike its predecessor, GPT-Image-2.5 has no discounted batch rate at launch. In early testing, Sunburst tends to cost more per image than Flare despite matching token prices, likely due to longer internal reasoning passes. OpenAI has not published an average price-per-image figure this time.

Editing consistency is the core claim

The central improvement OpenAI highlights is multi-step editing: the models are designed to modify only the requested elements of an image while leaving everything else — including previously edited details — untouched across several rounds of revision. OpenAI demonstrates this with a room redesign example where earlier edits persist unchanged through subsequent instructions, a known weak point of GPT-Image-2.

How ChatGPT routes users between Flare and Sunburst is not documented. API users can select a model explicitly, but the ChatGPT interface offers no visible toggle. Testing by the-decoder suggests the split roughly tracks ChatGPT's "Chat" versus "Work" surfaces, with Work sessions showing more consistent targeted edits and Chat sessions occasionally invoking the stronger model only at a "6 Pro" reasoning setting — behavior that may shift as the rollout continues.

New editing tools and leaderboard position

OpenAI is also adding a "Sketch" feature, activated with "@Sketch," letting users draw a visual template — a diagram, room layout, or poster outline — for the model to render. Additional features include ready-made prompt templates for formats like posters and logos, the ability to comment directly on images, and shareable prompts other users can reuse with their own inputs.

On Arena's text-to-image leaderboard, both new models currently hold the top two spots: Sunburst leads with a preliminary score of 1421 (about 3,100 votes), followed by Flare at 1399 (about 2,900 votes). GPT-Image-2 sits third at 1381 from roughly 78,700 votes — a far larger sample, meaning the new models' rankings could shift.

For provenance, OpenAI continues using C2PA metadata and is adding an invisible watermark via Google DeepMind's SynthID across ChatGPT, Codex, and the API. OpenAI says a single labeling method is insufficient, hence the layered approach.

GPT-Image-2.5 is available now worldwide to all ChatGPT tiers, including the free version, plus ChatGPT Work and Codex, across desktop, mobile, and web. OpenAI says weekly image generation volume across ChatGPT and the API now exceeds three billion images.

What this means

The headline claim — 50 percent lower latency with better quality — is significant if it holds up broadly, but the preliminary Arena scores rest on vote counts roughly 25 times smaller than GPT-Image-2's, so the leaderboard positions are not yet settled. More consequential for developers is the pricing structure: identical token rates mask real cost differences driven by the new "max" tier and the loss of batch discounts, meaning teams optimizing for cost need to benchmark actual token consumption rather than trust headline rates. The unclear routing logic in ChatGPT's consumer interface — where users can't explicitly choose Flare or Sunburst — also means quality may vary unpredictably by product surface until OpenAI documents or exposes that control.