GPT Image 2.5 is now live — two channels, pick the one that fits
Both GPT Image 2.5 tiers are now available on two channels: an official channel metered at 75% of list, and a flat-rate channel at $0.017 per image.
OpenAI released GPT Image 2.5 on September 8. Both tiers, Flare and Sunburst, are now available through two separate channels, so you can trade stability against cost on your own terms.
Official channel
Served through our most reliable upstream: stable throughput, predictable latency, failures are rare. Use it when a request has to go through.
Billing follows the creator's own token metering at 75% of their published rate, so what you pay tracks what you asked for. A low draft costs a fraction of a max render.
Model ids: gpt-image-2.5-flare-official, gpt-image-2.5-sunburst-official
Standard channel
A flat $0.017 per delivered image, whatever the quality tier or output size. A max render at 4K costs exactly what a low render at 1024x1024 costs.
Slightly less stable than the official channel; occasional retries are expected. In exchange, price stops being a variable you have to model.
Model ids: gpt-image-2.5-flare, gpt-image-2.5-sunburst
Which to choose
Both channels expose an identical parameter surface. Switching is one string in your request body.
The cost crossover sits between medium and high. Below it, metered billing is cheaper; at high and above the flat rate wins, and the gap widens as you climb.
Draft-heavy or low-tier work belongs on the official channel, where you also get better stability. Production work at high and above belongs on the standard channel, unless the request absolutely cannot fail.
What's new in 2.5
Two quality steps above the previous ceiling, xhigh and max. Transparent backgrounds fully supported rather than in preview. Edits accept input_fidelity to control how much source detail survives. Custom sizes accept any width and height that are multiples of 16, within a 1:3 to 3:1 ratio and up to 3840px on the longest edge.
Failed generations are refunded automatically.