Invest1 publisher2 min readPublished
OpenAI's new max image tier costs about 35 times its cheapest one
Flare and Sunburst share identical token rates of eight dollars in and thirty out per million, so what a picture costs is set by the quality tier and by however many tokens the model decides to spend.
The Investor · Invest desk
What happened
- OpenAI released two new image models under ChatGPT Images 2.5 and made both available to developers in the API.
- GPT-Image-2.5 Sunburst is aimed at more demanding visual work with tighter control over edits, and OpenAI says it needs longer generation times to deliver that.
- A 1024x1024 image costs about 0.006 dollars at the low tier, about 0.053 dollars at high, and roughly 0.21 dollars at the new max tier, which uses around 7,024 output tokens.
- OpenAI does not say, in the announcement or the documentation, when ChatGPT reaches for Flare and when it reaches for Sunburst.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- cost A team shipping a million 1024x1024 images a month is choosing between roughly 6,000 dollars at low, 53,000 at high and 210,000 at max. The quality tier is the budget line.
- decision Any workflow that depends on the stronger edit model now carries a hosting decision, because the model can be named in an API request but not chosen in ChatGPT.
- constraint Volume buyers who used the predecessor's batch discount have to re-plan around the standard rate, which makes the 0.6 cent low tier the cheapest route to scale for now.
- exposure A per-image budget is exposed to how long the model reasons: the-decoder.com's early tests had Sunburst costing more than Flare while both billed the same rate.
Multiply 7,024 output tokens by thirty dollars a million and you get 21.07 cents [1], the whole of the quoted price for a max-tier 1024x1024 image [9]. At that size the eight-dollar input rate barely registers [6]. Run the same rate backwards on the cheaper settings: 5.3 cents implies roughly 1,767 output tokens, and 0.6 cents about 200 [2]. Low and max are billed identically per token, and max spends about thirty-five times as many of them [3].
Compare that ceiling with the old one. According to the-decoder.com, the max tier of Images 2.5 costs the same as the high tier of GPT-Image-2 [10]. The old high therefore sat near 21 cents an image, and the new high, at 5.3 cents, costs about a quarter of that [4].
Flare and Sunburst share the rate card [6], and in early tests by the-decoder.com Sunburst usually cost more per image, which the publication puts down to longer reasoning runs [12]. A team that wants a per-image budget has to count tokens in its own logs. This time OpenAI did not publish an average price per image, and Images 2.5 has no cheaper batch rate [13][11].
OpenAI says more than three billion images a week are generated through ChatGPT Images and the GPT-Image models in the API [2], which works out to roughly 156 billion a year [7]. In tests by the-decoder.com, the dividing line runs between Chat and Work. In Work its prompts changed only what was asked, whatever the reasoning setting; in Chat other details kept moving in the follow-up images even with reasoning set to high [16]. Only at the "6 Pro" setting did the stronger model sometimes appear to engage [17].
OpenAI's latency figure is a ceiling, up to 50 percent below Images 2.0 [3], and the models are meant to preserve subjects from reference photos and follow editing instructions more reliably across several rounds [18]. If Flare holds most of that latency gain while beating GPT-Image-2 on quality [4], the default for most product teams is Flare at the high tier, or rather Flare at high until somebody counts the retries. Four attempts at 5.3 cents come to 21.2 cents, more than one max-tier image at 21 [6].
What to watch
- Whether OpenAI adds a batch rate to Images 2.5, as it had for the predecessor model.
- Whether OpenAI documents how ChatGPT picks between Flare and Sunburst for a given prompt.
- Independent latency measurements showing whether Flare lands near the 50 percent claim or well short of it.