Eight 0.5K drafts, then one 4K final: Nano Banana 2.1 costs $0.80
Iterate a hero image eight times at 0.5K for $0.60, then render the keeper at 4K for $0.20: $0.80 total on Nano Banana 2.1 through Sume, versus $1.60.

Iterating a hero image on Nano Banana 2.1 is cheaper when you draft small: eight 0.5K drafts at $0.075 cost $0.60, and one 4K final at $0.20 brings it to $0.80. Eight full 4K tries would cost 8 x $0.20 = $1.60.
Why the draft tier exists
The Sume catalog lists four resolution tiers for google/nano-banana-2.1: 0.5K (also written 512), 1K, 2K and 4K. Billing is per image at the tier you request, so the lower tier is a straight discount while you are still choosing a composition.
| Plan | Arithmetic | Cost |
|---|---|---|
| 8 drafts at 0.5K plus 1 final at 4K | 8 x 0.075 + 0.20 | $0.80 |
| 8 drafts at 1K plus 1 final at 4K | 8 x 0.10 + 0.20 | $1.00 |
| 8 tries at 4K | 8 x 0.20 | $1.60 |
| 12 drafts at 0.5K plus 1 final at 4K | 12 x 0.075 + 0.20 | $1.10 |
Does the final match the draft?
Not exactly. A new generation at a different tier is a new image, not an upscale of the draft. To keep the draft's composition, send it as a reference to an edit call at the higher tier, with a prompt that says to keep the layout. That adds one more 4K call but still lands near $0.80 plus the edit.
A practical limit
Small drafts hide fine detail, so they are poor for judging text. If label text is the point, draft at 1K. Remember that 4K calls are the most likely to return a 202 job envelope, so build the final step to handle that.
Where the savings come from
The saving against the all-4K plan is $1.60 minus $0.80, or $0.80, which is 50%. For 100 hero images the draft plan is 100 x $0.80 = $80.00 and the all-4K plan is 100 x $1.60 = $160.00, so the gap is $80.00.
The saving shrinks if you need more drafts per final. With 20 drafts at 0.5K, the plan is 20 x $0.075 + $0.20 = $1.70, which is already above eight 4K tries at $1.60. Set a draft cap before you start, and stop drafting when the composition is settled.
Pick the final tier by use. A web hero rarely needs 4K, so 2K at $0.15 is a fair final for it: 8 x $0.075 + $0.15 = $0.75.
Practical settings
Use a fixed seed-like discipline by keeping the prompt text identical between draft and final, changing only the tier. Sume does not accept a seed field (the docs say a request that sends one returns 400 unsupported_parameter), so you cannot reproduce a draft exactly. That is why the final should be treated as a new image, or generated as an edit of the chosen draft.
Name each draft file with the prompt hash and the tier so that you can find the one that you picked.
Mixed plans work well too. A common one is eight drafts at 0.5K, a 2K pass on the two best at $0.15 each, and a 4K final on the winner: 8 x 0.075 + 2 x 0.15 + 0.20 = $1.10. It costs $0.30 more than the simple plan and gives you a mid-size look at two finalists before you commit to 4K. Choose it when two compositions are close and the 4K call is the one you want to get right.
Sources
Related posts
More in Pricing
- Eleven v4 lists at $80 per million characters: is Turbo at $40 enough?
Eleven v4 and v3 list at $0.08 per 1,000 characters on the API, Turbo at $0.04. The sums for 1M characters, the Oct 12 discount, and where Sume TTS sits.
- Face swap beta: the catalog estimate is 15 seconds of rate, $2.76+
Avatar face swap beta is listed at the avatar video rate times the 15 second maximum source: $2.76 standard, $3.68 plus, $8.25 max.
- Game NPC barks: 800 lines at 60 characters, voice cost
800 NPC lines of 60 characters are 48,000 characters: $0.72 on MAI-Voice-2.1-Flash, $1.056 on MAI-Voice-2.1, $8.00 on Sume as 800 jobs.
- Gemini 3.8 Flash TTS prices rise 50% on Jan 1 2027: recompute
Gemini 3.8 Flash TTS lists $9.00 per million audio tokens until Dec 31, then rises 50%. The new Flash and Flash-Lite numbers, and how to compare with Sume TTS.
Written by Sume