Cheapest draft image on Sume: GPT Image 2.5 low at 0.74 cents

The cheapest way to draft an image on Sume is ChatGPT Image 2.5 at quality: low: $0.0074 for a 1024x1024 image, against $0.0250 for Grok Imagine, Qwen Image or Imagen 4 Fast and $0.0750 for Nano Banana 2 at its 0.5K tier. That makes the low tier 3.4 times cheaper than the next group and 10 times cheaper than Nano Banana 2's smallest size.

A draft only has to answer one question, whether the composition and subject are right, so the price of the throwaway pass matters more than the price of the final. The comparison below uses Sume's per-image prices, which are provider list prices times 1.25.

Draft prices side by side

Rows are ordered from cheapest. The last column marks whether the model accepts reference images, since a draft of an edit needs one. Prices for GPT Image 2.5 are Sume's quote at the stated size with a one-character prompt; the other rows are flat per-image prices.

What the comparison does not say

It does not say the cheapest row looks as good. low is the lowest of six quality tiers on ChatGPT Image 2.5, and OpenAI's image generation guide (read 2026-10-05) lists the tiers without scoring them. Judge the output on your own prompts.

It also compares unequal sizes. Nano Banana 2's 0.5K tier is a 512-pixel image, while the GPT row is 1024x1024, four times the pixels. Even so, the gap is not only the size: the Fal list rate for that tier is $0.06 before Sume's 1.25 factor, which is still eight times the GPT low-tier list rate.

A draft-then-final loop

The pattern that makes the cheap tier worth it is to draft several prompts at low, pick the winners, and send only those again at a higher tier. Same prompt, same size, higher quality:

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-image-2.5",
    "prompt": "Three-quarter view of a walnut desk lamp, warm light",
    "quality": "low",
    "image_size": "1024x1024"
  }'

The break-even

Ten low-quality drafts cost $0.0737, which is less than one Grok Imagine image at $0.0250 and less than a single medium GPT Image 2.5 image at $0.0165. Even if only one draft in ten survives to a high final at $0.0659, the loop costs $0.1396 per kept image, while ten direct high renders cost $0.6587.

The loop stops paying when nearly every draft survives. If you accept eight in ten, the savings shrink, and a single medium pass can be the better choice.

Edge cases to check

Imagen 4 Fast is text-to-image only on Sume, so it cannot draft an edit. Grok Imagine returns one image per call on Sume, so four takes means four calls. Pin image_size on GPT Image 2.5: leaving it out makes the quote reserve the upper bound of output tokens, which is far above the figure in the table.

Choosing the draft model

Draft with the model you will finish with. A cheap draft from a different model tells you about that model, not about the final one, so a Grok Imagine draft is a poor guide to a GPT Image 2.5 render even though it costs a quarter of a cent more than the low tier.

The exception is composition. If the question is only where things sit in the frame, any model on the table answers it, and the cheapest row wins. If the question is texture, lettering or lighting, draft on the final model at low and read the result with that limit in mind: lower tiers are expected to carry less fine detail.

Keep the draft size equal to the final size when you can. The size barely moves the GPT Image 2.5 price at low quality, so a smaller draft saves very little and hides cropping problems that only show at the final aspect ratio.