GPT Image 2.5 comes in two versions, Flare and Sunburst. Here's where each is available, how quality settings work, and what's still unclear.
What is GPT Image 2.5? #
GPT Image 2.5 is OpenAI’s updated image generation and editing model, released as a follow-up to GPT Image 2. It comes in two versions, Flare and Sunburst, and is available through ChatGPT, Codex, and the API. Flare is the faster, lighter model. Sunburst is the heavier, higher-quality version. OpenAI has not clearly stated which version runs by default inside ChatGPT or Codex, though hands-on testing suggests it’s Flare.
TL;DR #
- GPT Image 2.5 ships in two variants , Flare (fast) and Sunburst (heavy), rather than as a single model.
- ChatGPT and Codex users get access automatically , but OpenAI’s documentation doesn’t specify which variant runs by default in those surfaces.
- Testing points to Flare being the default inside ChatGPT and Codex, based on output speed and quality compared to what Sunburst is expected to deliver via the API.
- The API exposes both versions directly , letting developers choose Flare or Sunburst depending on whether speed or output quality matters more for their use case.
- Quality settings run on a sliding scale from low to max , giving another lever besides model choice to trade speed for fidelity.
- Default output resolution in ChatGPT and Codex comes in under 1080p (around 1672x941 in testing), so images generated there typically need an upscale for anything beyond casual use.
- The model’s real strength isn’t raw image quality, it’s conversational editing , especially when paired with a capable LLM that can direct multi-step edits and even do outside research to inform an image.
Other agents start typing. Remy starts asking. #
Scoping, trade-offs, edge cases — the real work. Before a line of code.
How is GPT Image 2.5 different from GPT Image 2? #
The core architecture is the same type of “thinking” image model that GPT Image 2 introduced: rather than relying purely on diffusion trained on image datasets, it reasons through a prompt’s literal instructions. That’s why it fills a wine glass all the way to the top when asked, or sets a clock to a specific time like 5:15, instead of defaulting to what looks aesthetically “correct” based on training data (diffusion models like Midjourney tend to default to 10:10 on clocks because it visually reads as a smiley face, and they underfill glasses because that’s what most training photos look like).
What’s changed with 2.5 is mostly refinement. The visible noise pattern that showed up in GPT Image 2 outputs has been reduced, though not eliminated. Composited or “pasted in” looking subjects still show up occasionally and need to be prompted out. The bigger jump isn’t in raw rendering fidelity, it’s in how well the model handles multi-step conversational edits, maintains character and scene consistency across generated angles, and responds to follow-up instructions without losing context from the original image.
Where can you access GPT Image 2.5? #
According to OpenAI’s documentation, GPT Image 2.5 is available to all ChatGPT, Codex, and API users. That’s a broad rollout by default, but the documentation stops short of specifying which of the two versions, Flare or Sunburst, actually powers ChatGPT and Codex out of the box.
Based on testing, the version running inside ChatGPT and Codex is most likely Flare, the smaller and faster of the two. That would track with typical product decisions: consumer-facing chat interfaces tend to default to the quicker model to keep response times reasonable, while the heavier Sunburst variant gets reserved for cases where a developer or user explicitly wants to trade speed for quality.
The API is where the choice becomes explicit. Developers building on top of GPT Image 2.5 can select Flare or Sunburst directly, rather than relying on whatever default a chat interface has baked in. This matters for anyone building a product on the API, since the two versions carry different tradeoffs in generation time and output polish.
How do the quality settings work? #
Separate from the Flare/Sunburst choice, GPT Image 2.5 offers a quality setting that ranges from low up to a maximum level. This gives another axis of control: even within a single version of the model, you can dial output quality up or down depending on whether you need a fast draft or a more refined final image.
Combined with the Flare/Sunburst split, this means there are effectively two levers to pull: which version of the model you’re running, and how much quality budget you’re asking it to spend on a given generation. For iterative work, like exploring compositions or running a wine-glass-and-clock style test, a lower quality setting on Flare is going to be the quickest path. For final output, especially anything meant for print or heavy cropping, Sunburst with a higher quality setting is the more sensible choice.
Seven tools to build an app. Or just Remy. #
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
What resolution do you actually get? #
Images generated inside ChatGPT and Codex came back at roughly 1672 by 941 pixels in testing. That’s below standard 1080p resolution, which means anything generated in those interfaces will likely need to be upscaled before it’s usable for print, large displays, or detailed cropping. This is a practical detail that matters more than raw model quality for a lot of real workflows: even a visually strong image is limited by that native output size unless you run it through an upscaling step afterward.
Is GPT Image 2.5 worth using over other image models? #
The honest answer is that it depends on what you’re optimizing for. Compared to diffusion-based tools like Midjourney, GPT Image 2.5 tends to produce results that are more literally correct (accurate clock times, properly filled glasses, requested details actually present) but sometimes blander or less aesthetically composed. Midjourney and similar diffusion models often produce prettier results at the cost of ignoring specific instructions.
Where GPT Image 2.5 clearly pulls ahead is in editing and iteration. Multi-turn conversational edits, like aging a photo forward by years, changing color grading to match a specific film reference, or generating consistent alternate camera angles of the same scene, are handled with a level of coherence that’s hard to get from a pure diffusion pipeline. Character consistency across generated angles was notably strong in testing, holding up better than some competing multi-image tools.
The model also benefits heavily from being paired with a capable reasoning LLM as a front end. When used through an agent that can browse for reference material, research real-world context, or self-critique its own output (like noticing cropped feet in a generated image and automatically regenerating with a wider frame), the image model becomes part of a larger creative loop rather than a one-shot generator. That combination, not the raw image model alone, is where a lot of the practical value shows up.
Frequently Asked Questions #
What’s the difference between Flare and Sunburst?
Flare is the faster, lighter version of GPT Image 2.5, and Sunburst is the heavier version built for higher output quality at the cost of speed. Both are exposed through the API, letting developers pick based on their priorities.
Which version does ChatGPT use by default?
OpenAI hasn’t officially confirmed this, but testing suggests ChatGPT and Codex default to Flare, the faster variant, rather than Sunburst.
Do I need to upscale images generated in ChatGPT?
Likely yes for anything beyond casual use. Native output in ChatGPT and Codex came in under 1080p in testing (around 1672x941), so images intended for print or large display benefit from an upscaling pass.
Is GPT Image 2.5 better than Midjourney?
It depends on the goal. GPT Image 2.5 follows literal instructions more accurately (correct clock times, filled glasses, specific requested details) while Midjourney tends to produce more conventionally polished aesthetics but with less prompt adherence.
Can I control image quality separately from choosing Flare or Sunburst?
Remy doesn't write the code. It manages the agents who do. #
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
Yes. GPT Image 2.5 offers quality settings ranging from low to a maximum level, independent of which model version you’re using, giving finer control over the speed-versus-fidelity tradeoff.