On September 8, 2026, OpenAI launched ChatGPT Images 2.5 and opened it to all ChatGPT and Codex users, while introducing two models to its API: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. The company says image generation latency is 50% lower than Images 2.0.
The Facts
According to OpenAI's official documentation, Images 2.5 improves on natural lighting, texture detail, preservation of the main subject from reference images, and adherence to multi-turn editing instructions. On the ChatGPT side, new features include @Sketch hand-drawn references, templates such as Poster and Merch, localized image annotations, and prompt sharing. On the API side, Flare is positioned as the default choice for high-concurrency workloads, while Sunburst targets premium workflows that require tighter control. Both are priced identically, at $30 per million tokens for image output.
How It Works
The dual-track design separates speed from precision. Flare emphasizes lower latency and suits creator content, product experiences, and high-volume generation; Sunburst devotes more compute to editing precision and takes longer to generate. The safety system card shows that Sunburst and Flare posted unsafe generation rates of 1.09% and 1.41% respectively in adversarial testing, below Images 2.0's 1.64%. That test set is not representative of production traffic, and OpenAI continues to maintain C2PA metadata and SynthID watermarks.
Industry Impact
Feedback from early users such as Adobe, Manus, and Higgsfield AI indicates the models have improved on transparent backgrounds and character consistency. At the same price, enterprises can choose Flare to speed up prototyping or Sunburst to safeguard finished quality depending on the scenario, but the dual track does not change the underlying token cost structure. Early adopters can use this to adjust internal pipelines, moving high-concurrency tasks to Flare while keeping Sunburst for premium needs.
Strategic Assessment
(The following is analysis, not fact.) The dual-track API formally offers a choice between speed and quality, but identical pricing means cost savings come mainly from reduced latency rather than a lower unit price. Commercial users still need to assess for themselves how Flare actually performs under specific precision requirements, and how cost-effective Sunburst is in high-concurrency scenarios. If differentiated pricing or public benchmark comparisons emerge later, the impact of this strategy will become clearer. For now, the available material shows only that OpenAI is attempting to address different workflow needs through model tiering; the actual results await market validation.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接