GPT-Image-2.5 Flare and Sunburst: OpenAI Splits Its Image API Into Speed and Precision

OpenAI's image API now has two flagship models instead of one. On September 8, 2026, the company introduced GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst — sibling models under the same 2.5 upgrade, split by workload rather than by generation. Flare is positioned as the fast, default choice for everyday and high-volume image generation; Sunburst is the slower, precision-first model for work where control over an edit matters more than turnaround. Both shipped in the Images API the same day as the consumer ChatGPT Images 2.5 rollout, and both extend the GPT-Image-2 lineage that OpenAI released in April and that the community had been tracking since a Codex leak surfaced on September 3.

For a developer the practical change is immediate: the decision is no longer "which image model do I call?" but "which of two siblings fits this workload?" The answer matters for cost and latency, because the two models are not interchangeable in speed. Everything below that is an OpenAI claim — speed, quality, and safety figures all come from the vendor and its announced partners, and none of the 2.5 models has an independent benchmark score yet as of September 9, 2026.

Two models, one upgrade

OpenAI's developer announcement frames the pair by use case. GPT-Image-2.5 Flare — the identifier is gpt-image-2.5-flare — is the default for most applications. OpenAI positions it for creator and social content, product experiences, visual search, rapid image prototyping, brand assets, and any high-volume generation pipeline. The company says Flare produces higher-quality images than GPT-Image-2 at up to 50 percent lower latency, and that in its own evaluations it generates roughly two to four times faster than GPT-Image-2; the announced early customer Manus has separately reported seeing 2–4× speedups with better transparent-background output in its own usage. The API reference tags Flare as the "small" model in the pair.

GPT-Image-2.5 Sunburst — identifier gpt-image-2.5-sunburst — is the premium half. OpenAI built it for workflows that benefit from tighter control across edits: production-ready campaign creative, polished product imagery, marketing collateral, and design tools where brand accuracy is the point. It takes longer per generation than Flare, and that extra compute buys what OpenAI describes as an extra level of precision and detail fidelity on edits. The API reference tags Sunburst as the "base" model — the larger, more deliberate of the two. Higgsfield AI, another announced early customer, says the model "understands what not to change," which is a useful shorthand for the whole 2.5 editing pitch: keep the parts of an image the instruction did not ask about intact, even in busy scenes.

Adobe says it is integrating both models into Firefly, which makes the pair the first OpenAI image models to reach that product line. Manus and Higgsfield AI are the other named early customers.

What actually improved

The release notes group the quality work under the same headline: sharper detail, stronger style adherence, and more control over edits. Underneath that, OpenAI lists specific changes that are worth reading carefully because they are the difference between a point release and a real step forward:

• More natural lighting and richer textures — the output-quality items that reviewers consistently describe as the visible jump over Images 2.0.

• Better subject preservation from reference photos — facial features and identifying details carry through a generation rather than being re-invented each time.

• More reliable multi-turn editing — earlier changes stay stable and each successive edit builds on prior work without degrading unrelated areas of the image.

• Up to 50 percent lower generation latency than Images 2.0 on the consumer side, which is the same speed story Flare tells on the API side.

• Sharper text and transparent-background handling, the two capabilities that image-API buyers benchmark most often; Manus's transparent-background note is the one independent-sounding corroboration so far.

OpenAI also says every output carries C2PA metadata and an invisible watermark, which is unchanged from GPT-Image-2's provenance approach and worth restating because it is a hard requirement in some regulated workflows.

What the API now exposes

The new models land in the same Images API endpoints developers already use — generation and edit calls — but the parameter surface has grown. The API reference for the 2.5 models shows a quality parameter with more headroom than GPT-Image-2 offered: auto (the default), low, medium, high, and two new tiers above high, xhigh and max. It also documents size options that run past the 2048-by-2048 ceiling GPT-Image-2 shipped with, up to 3840×2160 in landscape and 2160×3840 in portrait. The stated constraints are an edge length no larger than 3840 pixels, dimensions in multiples of 16, an aspect ratio no wider than 3:1, and total pixels between 655,360 and 8,294,400 — so the largest outputs sit just over 8 megapixels, in 4K-class territory. A background parameter accepts auto, opaque, or transparent, which is how transparent PNG output is requested.

Two caveats keep that paragraph honest. First, these are the settings documented for the 2.5 identifiers in OpenAI's reference; whether every tier and size is available on both models, and what each quality level costs in tokens, had not been fully documented at launch. Second, gpt-image-2 remains listed in the API — the 2.5 release is additive, and existing integrations do not break. The per-image price of the new models will depend on how many tokens each generation actually consumes, which means the old habit of budgeting by "dollars per 1024×1024 image" is the number to watch until real usage data accumulates.

Pricing hasn't moved

OpenAI kept the token rates identical to GPT-Image-2's: $8.00 per million image-input tokens and $30.00 per million image-output tokens, on top of the usual text tokens for the prompt. For context on what those rates meant on the previous model, Artificial Analysis quotes GPT-Image-2 (high) at about $211 per 1,000 images — roughly $0.21 for a high-quality 1024×1024 output — and the openai/gpt-image-2 page on OrcaRouter shows the same $8.00/$30.00 token split. Because the 2.5 models carry the same rates, the cost question is purely about token consumption per image, which is exactly what was not published at launch.

That is the pricing math that matters to anyone already calling gpt-image-2 through OrcaRouter. The OpenAI image models are among the 200-plus models behind the OrcaRouter API, and OrcaRouter passes provider list price straight through at 0% markup — the $8.00/$30.00 rates on the openai/gpt-image-2 model page are OpenAI's own numbers, not a reseller's. The day an upstream provider adds the gpt-image-2.5-flare or gpt-image-2.5-sunburst identifiers, the same API key calls them, and a vendor price change on either model is live on your side the same day it appears.

Flare or Sunburst: which one is for you

For most teams the honest answer is "start with Flare and escalate to Sunburst only when an edit has to survive scrutiny." The models split cleanly by workload rather than by raw quality, and OpenAI's own positioning says Flare is the default. The dimension-by-dimension read:

• Speed — Flare is the fast one: OpenAI claims up to 50 percent lower latency than GPT-Image-2, and 2–4× throughput in its evaluations and in Manus's usage. Sunburst deliberately trades speed for precision and takes longer per generation.

• Best workload — Flare for creator/social content, product experiences, visual search, rapid prototyping, and high-volume brand assets. Sunburst for production campaign creative, polished product photography, and design work where the edit must hold up.

• Editing behavior — Sunburst is the tighter editor; the "understands what not to change" feedback from Higgsfield applies to the 2.5 generation as a whole, but Sunburst is the model built for the most control across successive edits.

• API model tag — Flare is the "small" model in OpenAI's reference; Sunburst is the "base" model, meaning the larger and slower of the two.

• Price — identical token rates ($8.00 input / $30.00 output per million), so the cost difference is whatever extra tokens Sunburst's longer generations consume.

• Independent score — none yet for either model as of September 9, 2026; treat all quality and speed claims as vendor-reported until a leaderboard entry or a reproducible third-party test lands.

A reasonable pattern for a team already in production is to route the majority of traffic to Flare — it is the lower-latency default and the cheaper bet at the same token price — and to point only the edit-heavy, quality-critical slice at Sunburst. Because both models share the Images API shape and the same token rates, that split is a configuration change, not a migration. Automatic failover makes it safe to try before you commit: send a slice of real traffic to the new model and fall back to gpt-image-2 on an error or timeout, and your own prompts — not a launch page — decide which sibling earns the traffic.

Where the 2.5 launch lands in the image-model race

GPT-Image-2 was already the model to beat on the independent boards before this release. On the Artificial Analysis text-to-image leaderboard it holds the top spot at Elo 1178 — ahead of MAI-Image-2.6 at 1149, Nano Banana 2 at 1121, and Muse Image at 1116 — and on the image-editing board GPT-Image-2 (high) sits second at Elo 1117 behind Microsoft's MAI-Image-2.6 at 1122. The launch narrative for GPT-Image-2.5 is that OpenAI has taken that leading position and split it into a speed tier and a precision tier, which is a product strategy as much as a quality claim: rather than defending one flagship score, it is defending two price-performance positions at once.

The honest caveat is that none of this is measured yet for the new models. As of September 9, neither gpt-image-2.5-flare nor gpt-image-2.5-sunburst appears on the Artificial Analysis leaderboards, and no independent reproduction of OpenAI's speed claims has been published. The numbers above that do have independent backing — GPT-Image-2's Elo scores and per-image pricing — describe the predecessor, not the newcomers. On safety, OpenAI reports unsafe-generation rates of 1.41 percent for Flare and 1.09 percent for Sunburst against a 1.64 percent baseline on its internal evals; those are vendor figures, released to show the pair is safer than the model they extend, and they are worth treating as directional rather than definitive. The scale context comes from OpenAI too: it says users create more than three billion images a week across its platform.

What to watch next

Three things will tell you whether the split works better than the single flagship it extends. The first is independent latency and throughput numbers for Flare — OpenAI's 2–4× claim is the kind of figure a partner can verify in an afternoon, and Manus's corroboration is a start, not proof. The second is token consumption per image on both models, which will convert the unchanged $8/$30 token rates into a real per-image price; that data had not been published at launch. The third is where Sunburst lands on the image-editing boards, because that is the model's entire reason to exist: if its precision lead over GPT-Image-2 is real, it should show up against MAI-Image-2.6 and the rest of the editing top ten within a few weeks.

For now, the release is a genuine fork in the road for image-API buyers. GPT-Image-2.5 Flare is the sensible default migration path from GPT-Image-2 — same API, same token price, better speed on paper. GPT-Image-2.5 Sunburst is the one to hold out for if your work is edits under scrutiny. And if you are already calling gpt-image-2 through OrcaRouter, the switch to either sibling is a routing change on the same key, with the provider's list price passed through at 0% markup and failover available while you decide which model actually earns your traffic.