OpenAI splits its image model: Flare for speed, Sunburst for control
Two releases worth your attention this morning: OpenAI has turned its image model into a two-tier product line, and the former head of Huawei's Pangu model has shipped his first model out of a startup built on watching other models work.
OpenAI released GPT-Image-2.5 on September 8 — but the real change is that it shipped as two models instead of one. Flare is the fast default, which OpenAI says produces better images than GPT-Image-2 at up to half the latency; Sunburst is the slower, precision-focused sibling built for campaign creative and product imagery where an edit has to survive scrutiny. Both share the same token rate, so the cost difference is whatever extra tokens Sunburst's longer generations burn — which means nobody yet knows the real per-image price. Adobe says it will put both into Firefly, the first OpenAI image models to reach that product line, with Manus, Higgsfield AI and Runway also named as early customers.
The strategic read is more interesting than the quality claims. OpenAI's GPT-Image-2 already sits atop the Artificial Analysis text-to-image board, and rather than defending one flagship score it is now defending two price-performance positions at once: a cheap high-volume tier and a premium editing tier. That is a product decision dressed up as a model release, and it lands as image generation becomes a commodity that ships inside other people's software — hence the Adobe deal. Treat the speed and quality numbers as vendor-reported for now; neither model has an independent benchmark entry as of September 9.
TokenRhythm, the startup founded by former Huawei Noah's Ark Lab director Wang Yunhe, released NeoHorse-1 — a 4B and 9B "agent-native" model trained largely on the execution traces of other models. The company's open-source routing harness, OpenSquilla, decides which model handles each step of a multi-step agent task; those traces retain the routing choice, tool calls, environment feedback and failure paths, and were filtered for completeness and error recovery before post-training. Built with Infinigence-AI, Tsinghua, Peking University and Alibaba, the 4B version reportedly matches or slightly beats a 9B baseline overall. Parameter counts are public; price, context length and standard benchmark scores are not.
The bet is that a routing layer is a better training-data source than a static corpus, because it sees how agents actually fail. It is a clever dodge around the compute arms race, and the obvious risk is that a single stronger model eventually swallows the multi-model workflows the routing layer exists to optimize — though cost and controllability arguments keep that from happening quickly.
What to watch: whether Sunburst appears on the image-editing leaderboards within a few weeks, and whether TokenRhythm ever publishes comparable benchmark numbers for NeoHorse-1.
Which tier matters more to you in an image API — raw speed or edit precision? Tell us in the comments.
