OpenAI's GPT-Image-2.5 Sunburst tops Arena with 1421

OpenAI's release of ChatGPT Images 2.5 signals a split in its image generation strategy, offering two models tuned for different jobs rather than a single upgrade. The faster Flare targets general use with a 50 percent latency reduction, while the stronger Sunburst focuses on precise, multi-round editing—a capability that historically has been difficult for image models to master without altering unrelated parts of an image.

The company also introduced new quality tiers ("xhigh" and "max") that extend beyond the previous ceiling, but without a cheaper batch rate, which the predecessor offered. OpenAI notably declined to publish an average price-per-image figure, though early tests show Sunburst usually costs more per image than Flare despite identical token prices, likely because of longer reasoning runs.

While both models top the Arena leaderboard with preliminary scores, the practical rollout remains uneven. Our tests suggest ChatGPT's Chat mode currently uses the weaker model, with the stronger one appearing only at higher reasoning settings. As the rollout continues this week, the key question is whether OpenAI will give users explicit control or keep the model routing opaque, since the same token rates do not mean the same cost per image for end users.