Gemini API adds OpenAI-compatible Veo 3.1 video and image endpoints

• Google's Gemini API now exposes Veo 3.1 video generation and Gemini image models through its OpenAI-compatible surface, so existing OpenAI SDK clients can call them without a full client rewrite, according to the launch thread.
• The new mappings are concrete: Veo 3.1 goes through /v1/videos, while image generation uses images.generate for Nano Banana and related Gemini image models, as described in the launch thread and the docs post.
• Google is positioning this as a minimal migration path: both the announcement and the code example say teams can switch by changing "just 3 lines of code."

Google's update is narrow but useful for engineers already standardized on OpenAI client libraries. Phil Schmid's launch thread says Veo 3.1 and Gemini image models are now available through the OpenAI compatibility layer, with video requests routed through /v1/videos and image generation through images.generate. The linked compatibility docs frame this as "drop-in compatible" with the OpenAI Python and JavaScript SDKs.

That matters because this is not a new SDK story; it's a surface-area expansion inside an existing compatibility layer. Instead of adopting a Gemini-native client first, teams can keep the OpenAI package and swap the provider endpoint and model names, which the docs post says requires changing the API key, base URL, and selected model.

Google's code example shows the pattern directly: initialize OpenAI(...) with a Gemini API key, point base_url at Google's OpenAI-style endpoint, then call client.videos.create with model="veo-3.1-generate-preview". The sample polls client.videos.retrieve(response.id) until the job reaches completed, then returns a video URL.

The same migration pitch appears across the rollout. Logan Kilpatrick's supporting post repeats that Nano Banana and Veo work after updating three lines, while the docs summary adds that the compatibility layer also carries Gemini reasoning controls for supported models. For teams testing multimodal fallbacks or provider portability, the key implementation detail is that video and image generation now fit the same OpenAI-style client shape they may already have in production.