Meta Releases SAM 3.1... Offering API for Image and Video Object Detection and Tracking

Meta has released SAM 3.1, offering capabilities for object detection, segmentation, and tracking without hosting or tuning via the Meta Model API. This model runs on inference infrastructure specialized for the DETR architecture, ensuring high throughput from the start.

Detection, Segmentation, and Tracking in a Single Call

Unlike most multimodal models that provide only labels and bounding boxes, SAM 3.1 detects objects, generates precise segmentation masks, and tracks objects while maintaining their identity within videos in a single call. It operates in a zero-shot manner without requiring separate training data or fine-tuning.

Integration with Existing Pipelines and Pricing

SAM 3.1 shares the same endpoints, keys, documentation, and billing system as Meta's other models, allowing for easy integration into existing pipelines. The pricing is $2.50/1k per image and $0.20/1k per video frame. Users can execute their first request within 5 minutes via an OpenAI SDK-compatible client.