3 GPT low drafts, then a Nano Banana 2.1 2K final: 22 cents on Sume
Three GPT Image 2.5 low drafts plus one Nano Banana 2.1 2K final cost 22 cents on Sume ($0.07425 + $0.15). A cheap way to pick a composition first.
The short answer
Three GPT Image 2.5 low drafts plus one Nano Banana 2.1 final at 2K cost $0.22 on Sume: 3 x $0.02475 = $0.07425 for the drafts and $0.15 for the final, a total of $0.22425. A single 4K Nano Banana 2.1 image is $0.20 by comparison.
Mixing two models in one workflow is possible because both sit behind the same POST /v1/images call, and only the model field changes.
The arithmetic
The drafts settle the composition and the wording of the prompt. The final is then generated on the model you prefer for the finished look, with the winning draft attached as an input reference.
The table prices the plan against single-model alternatives for one finished 2K image.
| Plan | Arithmetic | Total |
|---|---|---|
| 3 GPT Image 2.5 low drafts + 1 Nano Banana 2.1 2K final | 3 x $0.02475 + $0.15 | $0.22 |
| 4 Nano Banana 2.1 2K images, keep one | 4 x $0.15 | $0.60 |
| 4 GPT Image 2.5 medium 2K images, keep one | 4 x $0.055625 | $0.22 |
| 1 GPT Image 2.5 high 4K image | 1 x $0.2225 | $0.22 |
| 8 Nano Banana 2.1 0.5K drafts + 1 final at 2K | 8 x $0.075 + $0.15 | $0.75 |
How to run it
Be honest about what the drafts tell you. A low-quality draft shows composition and subject, not fine detail, so judge only those.
Attach the winning draft as an input_references entry on the final call. Check the descriptor for the final model, since reference caps differ by model.
Retries and what to watch
The two-step plan only pays when the drafts remove retries. If the drafts do not change your choice, they are an extra $0.07425 per finished image.
These are the break-even points from the table.
- Against four Nano Banana 2.1 images at 2K ($0.60), the plan saves $0.37575 per finished image.
- Against four GPT Image 2.5 medium images at 2K ($0.2225), the plan costs the same to the cent.
- Against one GPT Image 2.5 high image at 4K ($0.2225), the plan costs the same, so choose it for the 2K look, not the price.
What the docs say to check
On GPT Image 2.5 the quality values are auto, low, medium, high, xhigh and max. If you omit quality the default is high, and auto reserves max, so pin the tier you budgeted for instead of leaving it to the default.
Read GET /v1/images/models before you pin a tier or ratio. A model accepts only the values its descriptors list, and the descriptors are the definitive source for what a call may send.
If you keep the final on GPT Image 2.5, then set quality to medium or high explicitly, since omitting it means high.
The numbers above are list prices read on 2026-10-08, not measurements. Sume figures are the catalog prices that already include its margin, Google figures are from its own pricing page, and quality or speed differences are not covered. Re-read both pages before you commit a large budget, because either side can change its rates.
Sources
Related posts
More in Use cases
- Cut a 3-minute Short from a 25-minute recording on Sume
Turn a 25-minute recording into a three-minute Short: detach audio, trim three beats, join them on Timeline with gapless audio parts. Calls and cost.
- A 3-minute YouTube Short from Omni clips: 18 takes, what it costs
YouTube says Shorts can run three minutes. At the 10-second Omni Flash cap that is 18 clips: the 720p and 1080p totals on Sume plus the Timeline join.
- A three-stage video funnel: Grok, MiniMax H3, Seedance 2.5 costs
Screen 20 ideas on Grok Imagine, keep 5 for MiniMax H3 at 768p, finish 1 on Seedance 2.5 at 1080p. All 5-second clips: the funnel costs about $10.23 on Sume.
- Thumbnail A/B set per video: three models for 31 cents on Sume
One thumbnail from each of GPT Image 2.5 high, Nano Banana Pro and Seedream 5.0 Lite costs 7 + 19 + 5 = 31 cents on Sume. Ten videos cost $3.10. How to run it.
Written by Sume