Meta turns SAM 3.1 into an API, segmenting a thousand images for just $2.5.

Beating AI News Flash: Meta has officially integrated its visual segmentation model SAM 3.1 into the Model API. Developers can input an image or video, then enter a phrase like "red bicycle" to directly locate the corresponding object and return bounding boxes and pixel-level masks. In videos, it can also track the same target all the way through, making it convenient for automatic blurring, adding special effects, and more.

Meta already released the open-weights version of SAM 3.1 in March this year. Compared with SAM 3, its biggest upgrade is Object Multiplex. Previously, multiple targets had to be processed one by one, but now up to 16 can be processed simultaneously at once. In Meta's tests, a single H100 increased its speed for processing multi-target video from 16 frames per second to 32 frames per second; when tracking 128 targets, the speed was about 7 times that of SAM 3.

Previously, developers needed to download the weights themselves, prepare GPUs, and set up a deployment environment. Now they can go directly through Meta's hosted API. Image segmentation costs $2.5 per 1,000 images, and video costs $0.2 per 1,000 frames. It can also be integrated directly using the OpenAI SDK.

However, the API version currently only accepts text prompts, while the open-weights version can also use visual prompts such as boxes and points to specify targets. The video API can track up to 16 objects per frame.