GPT Image 2 vs GPT Image 2.5: what changes when you migrate

To move from GPT Image 2 to 2.5 on Sume, change model from openai/gpt-image-2 to openai/gpt-image-2.5. Nothing forces you: the docs say ChatGPT Image 2 remains selectable, so existing requests keep working. What 2.5 adds is a wider set of options, and a request that sets an option the chosen model does not list is rejected with 400 unsupported_parameter.

The differences below are from Sume's Image API docs and catalog code, read 2026-09-29.

What is different between the two ids?

The two share the same aspect-ratio list. They differ in quality values, reference count and edit options.

Which requests break when I swap the id?

None of your old fields, because 2.5 lists every quality your GPT Image 2 requests could send. The break goes the other way: send xhigh, max or mask_url to openai/gpt-image-2 and you get 400 unsupported_parameter. That matters if you roll back, or if a config sends new options to both ids.

How do I move safely?

Read what each id accepts from the catalog, then switch one caller at a time. Keep the old id as a fallback in your config. If your code still sends the bare gpt-image-2 id, the docs say legacy bare ids are accepted as aliases for their org/slug equivalents, but prefer the full id in new code. Note that omitted quality defaults to high on 2.5.

curl "https://api.sume.com/v1/images/models" \
  -H "Authorization: Bearer $SUME_API_KEY"

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": "A red bicycle against a blue wall",
    "quality": "high"
  }'

Should I compare the two before switching?

Yes, on your own prompts. Sume's docs do not rank the models, and this post does not either. Run a set of prompts at high on both ids, which is the one quality both share, and judge the pairs before you change production. The blind-test script in GPT Image 2.5 arena: run your own blind test does exactly that.

What should I test after switching?

Replay a saved set of real requests against openai/gpt-image-2.5 and watch for three things: any 400 unsupported_parameter, any change in the shape of the output, and any change in cost. Because omitted quality defaults to high on 2.5, a request that relied on a lower default elsewhere may cost more, so set quality explicitly.

If you use references, note that the ceiling rises from 10 to 16 on 2.5 but your existing payloads still fit. Roll out to a small share of traffic first, keep the old id in configuration, and flip back if a result regresses for your use.