OpenAI shipped ChatGPT Images 2.5 on September 8, 2026. OpenAI's announcement framed it as "faster, sharper, smarter, with better tools for creating whatever you can dream of." Sam Altman posted his own gloss on X: "Images 2.5 is here. I don't think it can solve super difficult math problems, but it is really good and we hope you enjoy it."
That caveat is worth taking at face value — it tells you exactly where this release is strong (natural fidelity, targeted edits, speed) and where it still isn't (precise technical reasoning rendered as an image). This is the follow-up to ChatGPT Images 2.0 and gpt-image-2 from April 2026, and it lands into a market where Microsoft's MAI-Image-2.6-Flash, Google's Pics in Workspace, and Reve 2.1 have all shipped in the same three-month window.
TL;DR
| Question | Answer |
|---|---|
| Is this free in ChatGPT? | Same tiered access as before — free users get limited daily generations, Plus/Pro/Team/Enterprise get higher caps. No new 2.5-specific paywall announced. |
| What's new for developers (API)? | Two new models: GPT-Image-2.5 Flare (speed) and GPT-Image-2.5 Sunburst (precision), both reported topping image-gen leaderboards. |
| How is Flare different from Sunburst? | Flare trades detail for latency — good for drafts and high-volume, interactive loops. Sunburst trades speed for fidelity — good for final assets. |
| What's new for ChatGPT end users? | A Sketch tool to draw ideas before generating, comment-based editing on specific regions, and templates for posters/merch. |
| How much faster is it? | OpenAI claims 50% faster image generation than the prior version. |
| Can it still not do complex math/diagrams? | No — Altman said so directly. Treat technical diagrams as drafts, not finished figures. |
| What's the scale here? | OpenAI says over 3 billion images are generated weekly across its products. |
What actually changed
OpenAI's post lists four headline improvements, each aimed at a specific complaint from the ChatGPT Images 2.0 era:
- Faster generation — OpenAI claims roughly 50% faster than the prior version, which matters most for interactive workflows where a user is iterating in a chat loop rather than requesting a single final asset.
- Improved fidelity — more natural lighting, richer textures, and more physically accurate reflections, described as producing "more natural, recognizable images."
- Consistent details across edits — this is the practical fix for the single biggest complaint about earlier GPT-Image versions: asking for one small change (say, "make the sky darker") would often regenerate the entire image, shifting faces, product details, or composition that the user never asked to change. 2.5 is built to target exactly the region or detail specified, preserving everything else — including better face and pet likeness preservation across a sequence of edits.
- Comment-based editing — instead of writing a fresh prompt to describe a change, users can leave a comment on a specific part of the image and ChatGPT applies just that edit.
OpenAI also highlighted smoother multi-image sequences — stills that stay visually consistent frame to frame, enough that OpenAI Developer Experience engineer Charlie Guo posted: "GPT-Image 2.5 is good enough to do stop motion animation now!" That's a meaningful jump from earlier GPT-Image versions, where consistency across a sequence of related generations was one of the more common failure modes developers reported.
New tools in ChatGPT: Sketch and templates
Two additions target end users who aren't writing API code:
- Sketch — a drawing tool that lets a user rough out a composition (shapes, rough layout, placement) before generation, so the model has a spatial anchor instead of working from text description alone. This addresses the long-standing complaint that text prompts are a poor way to specify exact spatial relationships — the same limitation covered in how diffusion image generation works.
- Templates — preset starting points for common outputs like posters and merch designs, cutting the prompt-engineering step for people who don't want to write a detailed prompt from scratch. If you do want to write detailed prompts yourself, the community-maintained awesome-gpt-image-2 prompt library is still the most complete structured reference for GPT-Image prompt patterns.
GPT-Image-2.5 Flare vs Sunburst: which one to pick
The API-facing change is two new named models rather than one:
| Model | Optimized for | Pick it when |
|---|---|---|
| GPT-Image-2.5 Flare | Generation speed | High-volume batch jobs, interactive chat UIs, rapid draft/iterate loops where users see many candidates before committing |
| GPT-Image-2.5 Sunburst | Precision and fidelity | Final campaign assets, product photography stand-ins, anything where a visual error is expensive to catch after publish |
OpenAI says both models are "topping image-gen leaderboards" in their respective categories — Flare on speed-normalized benchmarks, Sunburst on quality-normalized ones. That's a now-familiar split: Microsoft ships MAI-Image-2.6-Flash alongside its full Pro-tier model on the same speed/quality trade, and the naming pattern ("Flash," "Flare," "Fast") has become a standard second SKU across image-model vendors in 2026. The practical build pattern hasn't changed either: generate previews on the fast model, then re-render only the approved candidate on the precision model — the same pattern developers already use for gpt-image-2 quality tiers.
If you're generating images inside a coding agent workflow rather than the ChatGPT app directly, the picks are similar to the pattern in generating images from Claude Code with the OpenAI skill: default to the fast tier while iterating on a prompt, and only spend the precision tier's cost once the prompt is locked.
Scale and what it signals
OpenAI disclosed that over 3 billion images are generated weekly across its products — a figure that puts image generation solidly among ChatGPT's highest-volume use cases, alongside text chat itself. That scale is the likely reason speed was the first headline improvement rather than an afterthought: at billions of generations a week, a 50% latency cut is a real infrastructure and cost story, not just a UX nicety.
What Altman's caveat actually means
Altman's post is unusually specific for a launch announcement: "I don't think it can solve super difficult math problems." Read literally, that's about image generation being used as a substitute for diagram-based reasoning — asking the model to draw out a geometry proof, a correct circuit diagram, or a precise chart with accurate data labels. Diffusion-based image models are still fundamentally pattern generators, not symbolic reasoners; they can produce something that looks like a correct diagram without the underlying relationships being correct. If your workflow needs a diagram where the content has to be exactly right — not just visually plausible — verify it manually or generate it with a deterministic charting tool instead, the same caveat that applied to every prior GPT-Image release.
What builders should do this week
- If you're on the API, test both Flare and Sunburst against your actual prompt mix rather than trusting the leaderboard claims directly — vendor-reported rankings vary by benchmark methodology.
- If you're editing existing images programmatically, re-test your edit prompts — the "targets only what you specify" claim, if it holds, may let you drop workarounds you built for the old "regenerates the whole image" behavior.
- If you're a ChatGPT end user doing design work, try Sketch before writing a long text prompt for anything with specific spatial layout requirements — posters, merch, and diagrams are exactly where a rough sketch beats a paragraph of description.
- Don't use it for diagrams that need to be factually correct. Altman said so himself.
Related reading
- How to Make Stop-Motion Animation With GPT-Image 2.5 — the step-by-step workflow behind Charlie Guo's viral stop-motion clip
- ChatGPT Images 2.0 and gpt-image-2 — the prior version this update follows
- GPT-Image-2 transparent backgrounds API preview — native alpha PNGs for campaign assets
- Awesome GPT-Image-2 prompt library — 530+ structured prompt templates and an agent skill
- MAI-Image-2.6-Flash launch — Microsoft's speed-tier competitor
- Google Pics: Workspace's AI image editor — Nano Banana 2-powered alternative
- Reve 2.1: 4K text-to-image arena leader — competing on the quality benchmark
- Generate images from Claude Code with the OpenAI skill — a coding-agent workflow that benefits from the Flare/Sunburst split
- How diffusion image generation works — the mechanics behind why comment-based edits are hard
- OpenAI — Introducing ChatGPT Images 2.5
Model names, benchmark rankings, and access limits are OpenAI's own claims as of the September 8, 2026 announcement. Re-check OpenAI's platform docs before building a production dependency on either Flare or Sunburst.