GPT Image 2.5 references: OpenAI's page says 4, Sume accepts 16

OpenAI's image-generation guide says 'up to four input images', while Sume's Image API docs say input_references takes a maximum of 16 on ChatGPT Image 2.5. Both statements are on their own pages, and they disagree. I could not reconcile them from the docs, so run your own test at the count you need.

The two numbers

Both numbers are quoted from the pages, not tested.

A test that settles it

Do not plan a 12-reference workflow from a doc line alone. Send 4, then 5, then your target count, and look at three things: the HTTP status, whether the output uses the late references, and usage.cost.

• Models: openai/gpt-image-2.5 or openai/gpt-image-2.5-sunburst.
• Read the input_references descriptor from GET /v1/images/models first; it is the source of truth for the catalog.
• Reference URLs must be publicly reachable; a bad URL returns a failure, not an output.
• Sume validates references against the catalog descriptor, so read the error body if a count is rejected.

Cost of more references

Sume's Flare and Sunburst rates are the same: $30 per million output image tokens, $8 per million input image tokens and $5 per million input text tokens. More references mean more input image tokens, so a 16-reference call costs more than a 1-reference call.