One config file for gpt-image ids before December 1

Sora showed what a hard-coded vendor id costs. Do not repeat it for images. OpenAI's deprecations page lists gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest with a 2026-12-01 shutdown and names gpt-image-2.5-sunburst or gpt-image-2.5-flare as replacements. Put image model ids in one config map keyed by your own tier names, then point the map at openai/gpt-image-2.5 or openai/gpt-image-2.5-sunburst on Sume, as the Image API docs list them.

A map that survives the next retirement

Your product should talk about tiers such as draft, standard and final. Only the map knows ids. When a vendor retires a model, you change one line and run a regression.

What to test before the date

The Image API docs say both 2.5 models support text-to-image, up to 16 image references, an optional mask_url, and background: auto|transparent|opaque. They accept quality of auto, low, medium, high, xhigh or max, and omitted quality defaults to high. image_size accepts presets, auto, or custom pixels where both edges are multiples of 16, within the documented pixel and ratio limits.

• Run your ten most important prompts on both models.
• Check transparent backgrounds if you used them.
• Check custom sizes against the multiple-of-16 rule.

Cost: read the page, then your usage

The same docs say Flare and Sunburst use the same Fal token rates: $30 per million output image tokens, $8 per million input image tokens and $5 per million input text tokens, with output estimates from OpenAI's size and quality calculator before input tokens and Sume pricing. Because auto quality reserves the max estimate, set an explicit quality if you want smaller reservations.

Do the same for video

Video needs the same discipline for a harder reason: the Sora models are already gone and OpenAI names no replacement. Keep a second map for GET /v1/videos/models ids. Image jobs follow the same job lifecycle described in Jobs and results.