gpt-image-2 can now output real transparent PNGs from one API flag
OpenAI's quiet late-August API drop also adds a prompt-cache dashboard, per-request region routing, and a temporary price cut on GPT-5.6 Sol.
Copy markdown
Transparent backgrounds, generated not stripped
Set `background: transparent` on gpt-image-2 and the API returns real alpha-channel PNGs (WebP too; JPEG can't hold alpha). Because the transparency is produced during generation rather than by a rembg pass, glass, hair and thin fibers come out clean — you can drop the background-removal step from your icon, sticker and product-shot pipeline. It's in preview, and OpenAI suggests leaving background descriptions out of the prompt.
A prompt-cache dashboard, finally
A new dashboard charts your cache hit rate over time, cache reads per write, and the cache-read / cache-write / uncached token split, filterable by model and service tier. If you run long system prompts, it's the first native way to confirm your caching is actually landing instead of quietly missing.
Route a single request to a region
Projects on Global geography can now pin an individual request to a processing region via a prefixed domain, instead of setting it once per project. Handy when one endpoint has to satisfy EU data-residency for some calls and lowest latency for others.
GPT-5.6 Sol gets cheaper until Nov 21
Sol drops to $4 per million input tokens and $20 per million output — a real per-call cut with a deadline. It stacks with the earlier Ultrafast tier that runs up to 14x faster, so latency-sensitive apps get both speed and a lower bill for now.