Podcast episode art for 50 episodes on GPT Image 2.5: medium vs high

Fifty square episode covers on ChatGPT Image 2.5 cost $0.83 at medium or $3.29 at high from text, as read from Sume's estimates.

5 min readSume
All posts

Fifty square episode covers on ChatGPT Image 2.5 cost about $0.83 at medium and $3.29 at high when generated from text at 1024x1024 on Sume. If each episode reuses your show's template image as a reference, the totals rise to $1.04 and $4.18.

These are estimates from the repository pricing code, with input tokens estimated. The decision is mostly about high, which costs about four times medium at this size.

The grid

Per image, rounded to four decimals. The reference column adds one estimated input image to keep a brand template fixed.

50 square covers at 1024x1024 on ChatGPT Image 2.5 (read 2026-10-07)
QualityPer cover50 coversPer cover with 1 reference50 with reference
low$0.0074$0.37$0.0094$0.47
medium$0.0165$0.83$0.0209$1.04
high$0.0659$3.29$0.0835$4.18

When medium is enough

Most podcast apps show a cover as a small tile, so fine detail is lost at the size listeners see it. Text is the exception: if the episode number or guest name must be legible, test it at medium first and move up only if letters fail.

The default on Sume is high when you omit quality, so a batch script that does not set it will bill the highest row in the table above. Set the field on purpose.

  • Generate a 4-up first with n to pick a direction, then reuse the winner as the reference.
  • Keep the prompt's text short; long titles are where small models drift.
  • Put the aspect ratio in the request as 1:1, not in the prompt.

Keeping fifty covers looking like one show

Consistency is the real cost in a series. The cheapest way to get it is a fixed template: one approved cover is passed as the reference in every call, and the prompt changes only the episode number, the guest and one accent element. The reference adds about $0.0044 per cover at medium, which is small next to the time you would otherwise spend fixing drift by hand.

Run a test of three covers at low before you commit. If the template survives three different titles, the batch of fifty is safe to run. If it does not, tighten the prompt rather than raising the quality tier, since high fixes detail but not layout.

  • Use one reference and one prompt skeleton for the whole season.
  • Generate a title-free background and add text in your design tool if lettering matters.
  • Archive the winning prompt with the season's assets.

A cheap loop

Draft the series look at low for under a cent a cover, lock the style, and run the 50 final episodes at medium. For the full five-quality ladder at six sizes, see the xhigh and max price grid. The Fal token rates behind the numbers are on the Fal model page.

If your show has a visual identity already, upload the existing cover as the reference and ask only for variations. This keeps the cost near the reference column in the grid and avoids a costly search for a new look. If the show is new, spend a little more on the first three covers at high, pick a direction you like, and then move the rest of the season to medium. The first covers set the tone for everything that follows, so they are the right place to spend.

Sources

Related posts

More in Use cases

All Use cases posts

Written by Sume