Thumbnail A/B test: five GPT Image 2.5 variants at medium vs high

Five 16:9 thumbnail variants at 1536x864 cost about $0.042 in GPT Image 2.5 output tokens at medium and $0.162 at high, before input tokens and Sume pricing. For twenty videos that is about $0.84 at medium for 100 candidates. The cheap way to run an A/B test is to generate all variants at medium, pick two, and rerun only those at high.

The budget

Per-image estimates come from the token formula behind Fal's GPT Image 2.5 Flare page and the Image API docs: $30 per million output tokens. 1536x864 is 16:9, both edges are multiples of 16, and it passes the size rules in OpenAI's image generation guide.

A workflow that spends little

Do not generate five variants at high. Generate them at medium, put them in front of whoever decides, and promote only the best two. A reference photo of the presenter's face is input tokens on top; at $8 per million per Fal's page it is a small line, but it is not zero, and the count is an estimate.

• Round 1: five variants at medium, one prompt per hook idea. The catalog lists n up to 4 for openai/gpt-image-2.5, so send n: 4 plus a second request of n: 1, or five single-image requests.
• Review: cut to two by a fixed rubric (face visible, three-word headline, contrast).
• Round 2: rerun the two prompts at high, n: 2 each.
• Ship: pick, then add any exact text in your editor if the model got a letter wrong.

Text on thumbnails

OpenAI's image generation guide says text rendering and composition precision remain areas for improvement, so plan to check every headline letter by eye. A variant with a misspelled word is a failed variant, however good the image looks. Keep the text to two or three words and consider adding it in a design tool after generation.

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": "YouTube thumbnail, shocked presenter on the left, bold headline NEW PHONE on the right",
    "aspect_ratio": "16:9",
    "quality": "medium",
    "n": 4
  }'

What to check

Read the n range on the model's catalog record before sending a count: the docs say per-model ceilings are lower than the route's 1 to 10, and the GPT Image 2.5 record lists 4, which is why the sample sends 4 and a fifth image goes in a second request. Large n is one of the configurations that tends to degrade to a 202 job, so handle that as described in Jobs and results. Also read usage.cost once to confirm the real billed amount before you scale to 100.