September 10, 2026. OpenAI released GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst to the API on 8 September, and its changelog says both "use GPT Image 2 token rates". The pricing page confirms it: $8 per million image input tokens, $2 cached, $30 per million output, identical to GPT-Image-2. If you produce creative at volume, do not read that as a flat bill. The rate is the price of a token, not of an image, and the tokens an image now costs are published nowhere.
What changed on 8 September
- Two API models. Flare is the default for most applications, and OpenAI says it delivers higher quality than GPT-Image-2 "at 50% lower latency". Sunburst targets precision editing and is explicitly the slower of the two, "with longer generation times".
- Two quality tiers above the old ceiling. Both models add xhigh and max. OpenAI's own guide states that earlier GPT Image models "support quality settings up to high".
- No token table. Both model pages carry the line "Token rates match GPT Image 2. The GPT Image 2 calculator does not estimate GPT Image 2.5 token consumption." The pricing page points readers to that same calculator.
- No Batch row. GPT-Image-2 appears in the published Batch pricing at $15 per million output tokens, half the standard rate. Neither 2.5 model does. OpenAI has not said Batch is unsupported, only that the rates are absent.
- One genuinely new capability. Custom sizes as a WIDTHxHEIGHT string, edges in multiples of 16, ratios between 1:3 and 3:1, no edge above 3840 pixels. Ad format ratios now map straight to a parameter.
Why the rate is not the price
Cost per image is the rate multiplied by the tokens an image consumes. OpenAI publishes the first number, withholds the second, and has added two tiers above the previous maximum. Higher quality consumes more tokens, which is why the guide's own cost advice is to "compare higher quality settings to find the right balance of detail, latency, and cost". Run the experiment and find out, in other words.
The contrast with the rest of the market is not theoretical. Google publishes the rate, the token count and the dollar figure together, stating that 1024 by 1024 outputs "consume 1120 tokens and are equivalent to $0.067 per image". Black Forest Labs prices FLUX per image outright, from about $0.014 on its smallest model. On OpenAI's published data the same number cannot be computed at all. Our arithmetic on the published $30 per million says every 1,000 tokens an image consumes costs three cents, so 1,000 images a month lands anywhere between roughly $34 and several hundred dollars, depending entirely on a figure OpenAI has not released.
The default is the exposure
Here is the part that costs money quietly. Both new models default to quality auto, which lets the model choose. On GPT-Image-2, auto could not reach above high, because high was the ceiling. Change the model string to 2.5 and leave every other parameter alone, and auto now selects from a menu containing two tiers above high, at a rate where a higher tier necessarily costs more. Nobody touched a pricing setting. In fairness to OpenAI, it does not publish what auto selects or how often, so this is an exposure rather than a measured increase. The fix is one line: pin quality explicitly before you migrate, then measure.
What it means for operators
If you generate a few images for a landing page, none of this matters. If you render catalogue imagery or cut ad creative in volume, three things follow. Pin the quality parameter instead of inheriting auto, and set it per workload rather than globally. Meter your first week against the actual invoice, because no reliable per image estimate exists yet for these models. And check whether your largest job needs to move at all: for an overnight, latency tolerant workload, GPT-Image-2 with its published Batch rate is the only image option on OpenAI's list with a halved output rate, which is an unusual conclusion to reach in an upgrade week. If those renders feed a storefront, that cost sits alongside the rest of the stack you already pay for, which is the lens we use in our Shopify app reviews, and building the render pipeline is AI automation work.
Frequently Asked Questions
Per token, no. The published rates are identical at $8 per million image input tokens, $2 cached and $30 per million image output tokens. Per image, possibly yes, because the new models add xhigh and max quality tiers above the previous high ceiling and higher quality consumes more tokens. OpenAI does not publish tokens per image for these models, so the per image figure cannot be calculated from published data.
OpenAI does not say. Both model pages state that the GPT Image 2 calculator does not estimate GPT Image 2.5 token consumption, and the pricing page refers readers back to that same calculator. The guide's own advice is to compare quality settings to find the right balance of detail, latency and cost, which is an instruction to measure it yourself. Meter a real workload against the invoice rather than trusting any published per image estimate.
They are two new quality levels available only on GPT-Image-2.5 Flare and Sunburst, sitting above the previous maximum. OpenAI's guide notes that earlier GPT Image models support quality settings up to high. Both new models default to quality auto, which means a straight model swap can let the model select a tier that was not previously reachable, without anyone changing a pricing setting.
Neither GPT-Image-2.5 model appears in OpenAI's published Batch pricing table, while GPT-Image-2 does at $15 per million output tokens, half the standard rate. OpenAI has not stated that Batch is unsupported for the new models, so treat this as missing pricing rather than a confirmed limitation. For a large overnight job the older model remains the only image option with a published halved output rate.