Izood MAG

AI — explainer

ChatGPT Images 2.5: sharper details, Sketch, and faster API options

5 min read
An enormous black-and-white photograph of an open ring-binder stands upright in the middle of the frame, scaled as though it were a wall.

ChatGPT Images 2.5: What Changed and Where to Get It

ChatGPT Images 2.5 arrives five months after Images 2 with sharper details, faster generation, more precise editing, and better sharing tools. Generation latency drops by up to 50% compared with Images 2.0. The release covers ChatGPT, Codex, and the API via two new developer models.

What ChatGPT Images 2.5 changes from Images 2

ChatGPT Images 2.5 is an enhanced version released five months after Images 2, which itself only arrived in April. The headline improvements: sharper details, faster generation, more precise editing, and better tools for creating and sharing. Generation latency is cut by up to 50% compared with Images 2.0.

For anyone editing real photos rather than generating from scratch, the meaningful changes are in fidelity and follow-through. Images 2.5 produces more natural lighting and richer textures, holds on to the subjects in your reference photos instead of drifting, and — the part I care about — follows editing instructions more reliably across multiple turns.

Skip this if you only generate one-off images and never revisit them. The upgrade pays off in the second and third edit, not the first prompt. If you're still deciding between models, the ChatGPT 4o vs 4 comparison covers the underlying model differences that affect image quality.

What these numbers don't tell you is how the multi-turn consistency holds up on longer sessions — that's where previous versions fell apart, and I'd test it before trusting the spec sheet.

New Sketch feature and creative tools

The new Sketch feature lets you draw directly inside ChatGPT and hand that drawing to the image model as a reference. You sketch the composition, the model fills in the rest. It's the fastest way to get a specific layout without wrestling with text prompts.

Skip this if you already generate images in a dedicated design tool — Sketch matters most when you want a rough idea turned into a finished image inside the same conversation.

OpenAI is also adding templates for popular image formats and prompt sharing. Templates give you a starting point for common aspect ratios and use cases. Prompt sharing means you can pass a working prompt to someone else instead of describing what worked.

The catch: none of this is useful until the model interprets your sketch the way you intended. In practice, you'll iterate more than once.

GPT-Image-2.5 Flare vs Sunburst: which API model to choose

Developers get two versions of Images 2.5: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. The decision comes down to one question: do you need speed, or do you need editing control across a sequence of changes?

Flare is the workhorse. It delivers higher-quality images than GPT-Image-2 at 50% lower latency, and it's the default choice for most applications. If you're generating images for user-facing features or batch processing — anything where a user waits on a spinner — pick Flare and move on.

Sunburst is the specialist. It's built for premium visual workflows that benefit from tighter control across edits. Think iterative design work where you're refining the same image through ten revisions and every adjustment needs to respect the previous one.

My honest read: if you're not sure which one you need, you need Flare. Sunburst earns its place only when the edit chain — not the single generation — is the product. Skip it if your workflow is generate-and-serve; the tighter edit control doesn't pay for itself at that volume.

Codex integration and native image generation support

Codex subscriptions now include native support for ChatGPT Images 2.5 — no separate image API integration required. If you're not a Codex subscriber, skip this section; everything else in the guide works through the standard API.

The change matters because of how it worked before. To generate images with native tools, you had to route requests through an LLM, which would interpret your prompt and call the image tool on your behalf. That added latency and a layer of unpredictability: what you asked for was filtered through what the model thought you meant.

Now you can specify the image provider directly. The supported provider model is chatgpt:gpt-image-2.5-sunburst, which points Codex at the Sunburst variant specifically. Flare isn't exposed through this path at the time of writing. If your workflow needs Flare's characteristics, you're still using the LLM-routed approach or a separate API call.

API details: quality levels, backgrounds, aspect ratios, and output

The API exposes five quality levels — low, medium, high, xhigh, and max. Cost scales with fidelity, so pick the lowest tier that still holds up for the use case. A thumbnail in a feed does not need xhigh, and paying for it there is waste.

Backgrounds are binary: transparent or opaque. No gradient presets, no scene options — if you want a product shot on a textured surface, you generate the surface or composite it later. For anything destined for a sprite sheet or layered mockup, transparent is the setting I actually leave on.

You control composition with aspect ratios like 16:9 and 1:1, and resolution with sizes including 2K and 4K. Skip the 4K tier unless you're printing or cropping hard — the file sizes get obnoxious in a CI pipeline and most screens will never show the difference.

Output lands wherever you point the --output flag. No surprises, which is the quiet part that makes batch generation scriptable; you can loop over a prompt list and not babysit a download dialog.

Rollout: where to get Images 2.5 today

Images 2.5 is rolling out today to ChatGPT, ChatGPT Work, and Codex users across all tiers on desktop, mobile, and web. No waitlist, no separate app — if you have any of those three products, you have the model. The release follows last week's GPT-6 Astra launch, so OpenAI is shipping image and text updates on roughly a weekly cadence right now.

Skip this if you're still on the free tier of the old standalone ChatGPT app — but if you've hit login or rollout issues in the past, this one doesn't require a fresh install. Just open the existing app or site and start a new image request.

How Images 2.5 compares to the previous Images 2 release

ChatGPT Images 2 arrived in April with a meaningful hardware-style bump: 2K resolution output and multiple aspect ratio support, plus the ability to pull from web research for current information. It also picked up non-Latin language expertise and two intelligence levels: Instant and Thinking. That was the foundation.

Five months later, Images 2.5 is the refinement pass. Skip this comparison if you never touched Images 2 — the sections above give you the full picture. For upgraders, the delta is narrower than the version number implies.

Both releases share the same core output specifications. The changes in 2.5 concern how precisely the model edits and how sharply it renders detail, not the resolution ceiling or the aspect ratio menu you already know. Think of 2.5 as a lens grind on an existing sensor, not a new sensor.