Meta launches Segment Anything Model (SAM) 3.1 as an API, with segmentation of 1,000 images costing just $2.5.

Beating AI Express: Meta officially integrates visual segmentation model SAM 3.1 into its Model API. Developers can input images or videos, then enter a phrase like "red bicycle" to directly locate the corresponding object, returning bounding boxes and pixel-level masks. The feature also enables continuous tracking of the same target in videos, supporting use cases such as automatic blurring and special effects. Meta released the open-weight version of SAM 3.1 back in March this year. Its biggest upgrade over SAM 3 is Object Multiplex: previously, multiple targets had to be processed one by one, while up to 16 can now be handled simultaneously in a single pass. In Meta’s tests, an H100’s processing speed for multi-target videos rose from 16 frames per second (FPS) to 32 FPS; when tracking 128 targets, the speed is roughly 7 times that of SAM 3. Previously, developers had to download weights, prepare GPUs, and set up deployment environments on their own; now they can directly use Meta’s hosted API. Image segmentation costs $2.5 per 1,000 images, and video processing costs $0.2 per 1,000 frames. The API can also be directly integrated via the OpenAI SDK. However, the API version currently only accepts text prompts, while the open-weight version supports additional visual prompts like bounding boxes and points to specify targets. The video API supports tracking up to 16 objects per frame.