If you generate section images for a PDF guide, you have a choice between two Stability models: Stable Image Core and Stable Image Ultra. We used both for the same job, pdf_guide.section, across 292 calls in the last 90 days.
Core handled 191 of those calls and Ultra handled 101. Neither came back with a failure on our side. The differences that show up in our figures are cost, speed and which one is still in use.
This post goes through each figure for each model, says which one we kept, and says plainly what the numbers cannot tell you.
What each model cost us per call
Stable Image Core (stability/stable-image-core) cost 0.03 USD per call. Across its 191 calls that came to 5.73 USD in total.
Stable Image Ultra (stability/stable-image-ultra) cost 0.08 USD per call. Across its 101 calls that came to 8.08 USD in total.
Ultra cost 2.67 times what Core cost per call, which is 0.05 USD more on every call. Core was the cheapest of the two per call.
These are the amounts we paid on our own volume. They are not a vendor rate card, and we are not claiming anything about Stability's published pricing.
How long each one took
Core had a median of 8.7 seconds per call. Ultra had a median of 15.2 seconds.
Ultra was 1.75 times slower than Core, or 6.5 seconds slower on the median call. Core was the faster of the two.
The figures we hold are medians. We have nothing here on the slowest calls for either model, so we are not describing the worst case.
Failures
Neither model produced a failure on our side. Core had 0 failures in 191 calls and Ultra had 0 failures in 101 calls, so both failure rates were 0.0 percent.
That is a count of calls that came back as failures for us. It is not a statement about Stability's uptime.
Which one we kept and which we stopped using
Core is the one still in use. It was first used on 2026-07-20 and its last call in the window was on 2026-10-09.
Ultra ran from 2026-07-15 to 2026-09-09 and has not been used since. It was the first of the two to appear, and Core's first call came five days after Ultra's.
The record does not say why we stopped using Ultra. The facts we hold are the call counts, the costs, the timings and these dates, and none of them states a reason.
What the token fields tell you here
They tell us nothing here. Billed input, cached input, total input and output tokens per call are all zero for both Core and Ultra, and the cached input percentage is empty.
Billing for these two image models is a flat amount per call in our figures, with no token breakdown to read. So the gap in cost per call between them is not something we can trace to input size or to cache hits. All we can say is what each one cost us per call, which is 0.03 USD for Core and 0.08 USD for Ultra.
What these numbers do not tell you
These are our prompts, our job and our volumes, measured over one 90-day window. A reader with a different job may see different numbers. This is not a benchmark.
The two models also did not have equal exposure. Core had 191 calls and Ultra had 101, and they were in use over different stretches of the window. Ultra's last call was on 2026-09-09, while Core kept going to 2026-10-09.
Our figures contain no measure of image quality. Nothing here says whether Ultra's output was better or worse than Core's for a PDF guide section. If quality matters to your decision, these numbers do not answer it.
They also say nothing about other jobs. Everything above is about pdf_guide.section and only that.
On our figures, pick Core for this kind of job: it cost 0.03 USD per call against 0.08 USD, took a median of 8.7 seconds against 15.2, and had no failures in either case. Core is also the one we still use. We would pick Ultra only if looking at the images side by side shows a difference worth 2.67 times the cost and 6.5 more seconds per call, because our figures do not measure that.
Originally published at neuragrowth.co. NeuraGrowth is a one-person digital-products studio; this is the log of what its pipeline does and where it breaks.
