ChatGPT Images 2.5: Flare vs Sunburst and the creative upgrade that matters
OpenAI split its latest image model into two API flavors — a fast Flare and a precise Sunburst — alongside Sketch, templates, and comment-based editing. Here's how the split works, what it costs, and which model fits which job.
On September 8, 2026, OpenAI released ChatGPT Images 2.5 — and it did something unusual for a model launch. Alongside the consumer upgrade, it split the developer-facing API into two distinct models: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. One is built for speed and volume; the other for precision and control.
It's a notable shift in how OpenAI packages image generation. Instead of asking everyone to use one model and tune knobs like quality tiers, the lab is now asking you to pick a lane. That choice is worth understanding, because the wrong pick either wastes money and patience — or ships worse work than your tools can deliver.
What Images 2.5 actually improves#
Before the split, the headline upgrades over Images 2.0. OpenAI says Images 2.5 cuts generation latency by up to 50%, improves instruction-following across multiple editing rounds, and preserves subjects from reference photos more reliably — more natural lighting, richer textures, faces and details that carry through edits instead of drifting.
That last point is the one creative teams will feel most. Earlier versions had a bad habit: ask it to change a jacket's color, then adjust the background, then tweak the lighting, and somewhere along the way a face would quietly reshape or a logo would vanish. OpenAI says 2.5 keeps earlier changes consistent across multi-turn edits without degrading quality. If you produce product shots or brand assets rather than one-off art, this is the upgrade that matters.
OpenAI also notes that more than 3 billion images are created every week across ChatGPT Images and the GPT-Image API models — context for why shaving latency and stabilizing multi-turn edits is a big deal at that scale.
Four new tools in ChatGPT#
The model upgrade ships with four new product features in the ChatGPT app:
- Sketch — draw directly inside ChatGPT (type
@Sketch) and use your rough drawing as a visual reference for the final image. Room layouts, outfit concepts, composition ideas — you no longer need words for everything. - Templates — starting points for popular formats like Poster and Merch, so you begin from a structure instead of a blank canvas.
- Comment-based editing — pin a comment on a spot in the generated image for a focused fix, instead of rewriting the whole prompt.
- Prompt sharing — share an image together with the prompt that made it, so someone else can adapt the idea with their own photos and details.
These roll out to all ChatGPT, ChatGPT Work, and Codex users across tiers on desktop, mobile, and web. (Coverage notes that Sketch is available on mobile, while Templates don't appear in the Work surface, and existing generation limits are unchanged — so expect some tier-to-tier quirks.)
Flare vs Sunburst: the actual split#
Here's the core of the launch. In the API, OpenAI now offers two models:
| GPT-Image-2.5 Flare | GPT-Image-2.5 Sunburst | |
|---|---|---|
| Role | Default for most applications | Premium lane for precision work |
| Strength | Speed, throughput, everyday quality | Tightest editing control |
| Latency | Up to 50% lower than GPT-Image-2; OpenAI's evaluations show 2–4× faster throughput | Longer generation times by design |
| Quality | Higher quality than GPT-Image-2 | Built for detailed creative work at production standards |
| Provenance | C2PA metadata + invisible watermarking | C2PA metadata + invisible watermarking |
| Best for | Social and creator content, product experiences, visual search, rapid prototyping, high-volume generation | Production-ready campaign creative, polished product imagery, multi-round refinement |
A few important nuances:
- Flare is the default, not the dumbed-down tier. OpenAI describes it as bringing "the same improvements in quality, editing, and speed" to the API. Both models get the editing improvements; Flare isn't a no-edit budget option.
- Sunburst trades time for control. It's aimed at workflows "where editing precision matters most" — the extra level of control across edits that campaign creative and print-quality product imagery need.
- The model cards mark Batch as unsupported for 2.5, per launch coverage — so if your pipeline leans on batch processing, plan around that.
What it costs (and what to watch)#
The API pricing is straightforward at the token level: $5 per 1M text input tokens, $8 per 1M image input tokens, and $30 per 1M image output tokens — identical token rates for both Flare and Sunburst, and the same rates GPT Image 2 charged. OpenAI points developers to its pricing page for details.
The honest caveat: OpenAI's pricing calculator for GPT Image 2 does not estimate GPT Image 2.5 usage, per launch coverage — so per-image costs are unverified as of this writing. One source estimates roughly $0.006 to $0.211 per 1024×1024 image depending on quality tier, and notes the high tier now uses the former medium token budget. Treat those as estimates, not official figures, until usage data lands.
Also of note from the API guide: supported sizes include standard options (1024×1024, 1536×1024, 1024×1536) plus custom dimensions up to 3840 px on the long edge, quality levels from low through max plus auto, transparent or opaque backgrounds, and PNG/JPEG/WebP output.
Which one belongs in your workflow?#
This is the practical question the split raises. Here's a decision framework:
Pick Flare when:
- You're generating at volume — social content, e-commerce listings, marketing variations.
- Speed matters more than pixel-level fidelity: rapid prototyping, mood boards, storyboards, visual search results.
- You're iterating in conversation and want the edit loop to feel instant.
- You're migrating from
gpt-image-2and want the closest drop-in: OpenAI frames Flare as the default for most applications.
Pick Sunburst when:
- The asset has to survive scrutiny — campaign creative, hero imagery, print work.
- You're doing multi-round refinement where small errors compound, like layered retouching on product photos.
- A precise brief has to land exactly: composition, copy, brand treatment preserved while one element changes.
If you're only using ChatGPT (not the API): you don't choose — the app runs Images 2.5 as a single experience. The Sketch, templates, and comment-editing features are where your workflow changes.
If you're building with the API: start with Flare as your default, route the precision lane to Sunburst, and don't rewrite your cost model until you have real usage data. The token rates match GPT Image 2, but actual consumption per image for 2.5 is still unverified.
Safety and provenance#
Both models carry the same safeguards: C2PA metadata and invisible watermarking to identify AI-generated images. OpenAI cites safety metrics of 1.09% unsafe generations for Sunburst and 1.41% for Flare versus a 1.64% baseline. Launch coverage also reports Adobe is integrating both models into Adobe Firefly — the first OpenAI image models on that platform — though details remain thin as of publication.
Takeaway#
The Flare/Sunburst split is OpenAI admitting what practitioners already knew: one image model can't be both the fastest thing in the pipeline and the most careful. Flare is the workhorse — faster, cheaper to iterate with, and the right default for most applications. Sunburst is the craft tool — slower, tighter, built for the final 10% of polish that separates drafts from deliverables. Most teams will end up using both: Flare to explore, Sunburst to finish.