OpenAI's GPT Images 2.5 Splits into Two Models—but Pricing and Speed Trade-Offs Remain Murky

OpenAI has introduced two new image-generation models this week, Flare and Sunburst, each designed to handle selective edits—changing specific elements of an image without altering the rest. Yet the company has left developers with more questions than answers about which to choose.

Two Flavors, One Unclear Trade-Off#

Flare positions itself as the faster, mainstream option: OpenAI calls it "the default choice for most applications" and claims it delivers higher-quality output than its predecessor GPT-Image-2 while cutting latency by half. That makes it suitable for rapid workflows like social media content creation, visual search, and quick prototyping.

Sunburst takes the opposite approach, targeting premium creative work. OpenAI describes it as built for "tighter control across edits" and production-ready campaigns where precision matters—think high-stakes product imagery or polished campaign assets. The catch: Sunburst is slower, though OpenAI refuses to say how much slower.

Both models carry identical published token rates ($5 per million text input tokens, $8 per million image input tokens, $30 per million image output tokens), yet that transparency masks a deeper problem: the company provides no way to estimate how many tokens each model will actually consume for the same task.

The Pricing Puzzle#

Because image generation fees depend on token consumption, and OpenAI has not disclosed how many tokens each model uses per image, developers cannot predict whether Flare or Sunburst will cost more for identical work. Making matters worse, OpenAI explicitly states that "the GPT Image 2 calculator does not estimate GPT Image 2.5 token consumption"—so the old cost-estimation tool won't help.

The only real solution is trial and error: developers must experiment with both models in production to discover which offers better value for their use case.

Better Editing, But at What Price?#

Beyond the model choice itself, GPT Images 2.5 does promise genuine improvements. The new generation can edit individual elements—a product, background, or text—while keeping the surrounding scene intact, a task that previously often required rebuilding entire assets. The models also follow multi-step editing instructions more reliably and understand complex visual prompts more accurately than before.

For teams managing high-volume image workflows, these refinements could reduce iteration cycles and save time. Whether Sunburst's added precision justifies its longer processing time remains an open question—one each developer will need to answer by testing it themselves.