Meta Adds SAM 3.1 to Its API for Object Detection, Segmentation and Tracking

Meta has released SAM 3.1 through its Meta Model API, giving developers hosted access to a vision model that can identify objects, draw pixel-level boundaries around them and follow them through video. The release packages tasks often handled as separate outputs into one request, with Meta positioning it as a zero-shot system that needs no training data or fine-tuning.

The launch makes the model available as a self-serve service. Meta offers API access, documentation and a playground for testing media before code is written. It also says developers can point existing OpenAI SDK-compatible clients at the Meta Model API, which could make the service easier to adopt for teams already using those clients.

A text prompt becomes a detailed track

SAM 3.1 is designed to detect objects in images and video, return a bounding box and a segmentation mask for each match, then track objects over time while preserving their identities. A segmentation mask marks the exact pixels belonging to an object, rather than only drawing a rectangular box around it. Meta says a developer can name an object with a short text prompt and receive those results without adding separate models.

What Meta says the service returns

• Object detection, locating matching objects in an image or video.
• Pixel-level masks that outline every matching object.
• Video tracks intended to retain an object’s identity from frame to frame.

$2.50 per 1,000 imagesImages

Meta lists SAM 3.1 image segmentation at $2.50 per 1,000 images.

$0.20 per 1,000 framesVideo

Meta lists video use at $0.20 per 1,000 frames.

The service layer is part of the pitch

Meta says SAM 3.1 runs on inference infrastructure purpose-built for the model’s DETR architecture, and promotes production throughput without customers needing to host or tune the model themselves. Those performance and deployment claims come from Meta, but they clarify the product’s target: teams that want a ready-made vision capability rather than a model to operate.

Reliability remains the practical test

The next question is whether the zero-shot approach holds on the media and object categories a developer actually needs to handle. Meta has published access routes and per-image and per-frame rates, but reliable results on particular imagery and video will still determine how useful the service is in production.