OpenAI has added GPT Image 2.5 Sunburst and GPT Image 2.5 Flare to its API pricing page, and the notable thing is the number.
The prices
Both variants are listed at $8 per million image input tokens and $30 per million image output tokens. Cached image input is $2. Text input is $5 per million, with cached text at $1.25.
Set against the older models in the same family, the direction is unusual. GPT Image 1.5 costs $8 in and $32 out. GPT Image 1 costs $10 in and $40 out. The small GPT Image 1-mini remains the cheapest at $2.50 and $8.
Why this is backwards
A new generation normally arrives at a premium and pushes the previous one down. Here the newest models are priced below their predecessors on output, which is where image generation costs actually land.
The likely explanation is competitive rather than technical. Image generation is the most crowded category in the market, with strong open-weight options and a long list of vendors distributing models through aggregators. Charging more than last year's model is difficult when a user can switch providers inside the same interface.
What it means in practice
If you are still calling GPT Image 1 or 1.5 in production, you are paying more for an older model. That is a rare situation and an easy thing to check.
The wider lesson is about how quickly this line moves. Pricing for image models has become genuinely volatile, so a cost estimate built six months ago is unlikely to still be accurate. Re-check the pricing page before committing to a provider, and prefer platforms that let you swap the underlying model without rebuilding your pipeline.