Computer vision without the computer vision team.

Yes. New unbilled accounts receive 100 free image tokens with no card required. The one-time 14-day trial starts with the first accepted job. It supports up to three images per job and downloads; video requires paid access. Only successfully processed images consume tokens, and credits do not reset monthly. See the trial guide.

SAM 3 is Meta’s Segment Anything Model 3, released in November 2025. Describe a concept such as “person” or “red car” and it detects, segments, and tracks every match across images and video. It also supports point, box, and exemplar prompts.

The easiest way to use SAM 3 is SegmentationAPI. Instead of provisioning GPUs, managing model upgrades, and building your own queueing and storage, you send a single REST request and get masks back. SegmentationAPI is the easiest and most production-ready way to run SAM 3 on the market: managed infrastructure, a 99.99% uptime SLA, and simple token-based pricing with no setup required. See the quickstart to try it yourself.

SegmentationAPI runs SAM 3 through a hosted REST API. Use it to auto-label datasets, build annotation tools, or segment video. It returns masks, confidence metadata, and vector outputs without GPU management. See the API docs to get started.

SAM 1 and SAM 2 use points or boxes, and SAM 2 added video tracking. SAM 3 adds open-vocabulary text prompts, built-in counting, and unified image and video tracking. Meta reports roughly a 2× gain on promptable concept segmentation over prior systems. See our SAM 3 alternatives comparison.

Each token costs $0.02 and covers one processed image or video frame. There are no seat fees or GPU commitments. Use the pricing calculator to estimate your spend.

Yes. Give SAM 3 one text prompt and it tracks matching objects across the video without per-frame annotation. SegmentationAPI bills one token per processed frame.

Concept segmentation uses a short noun phrase instead of a click. The model returns a mask for every matching object. Open-vocabulary means the phrase is not limited to a fixed class list.

No. SAM 3 is zero-shot: send an image or video with a text, point, or box prompt and get masks back immediately, with no dataset collection, training, or fine-tuning required. Its zero-shot mask quality is high enough to use directly as training labels for downstream models.

Self-host if you need on-prem deployment and can operate the GPUs, queues, storage, and scaling. Use the API when you want managed infrastructure, model upgrades, retained assets, and previews.

SAM 3 leads open-vocabulary segmentation. On the SA-Co Gold concept-segmentation benchmark it scores 53.9 cgF1 versus 24.6 for OWLv2 and 13.0 for Gemini 2.5 Pro, and it tops counting accuracy on CountBench at 93.8%. For real-time edge deployment, lighter models like YOLO or RF-DETR are faster; a common pattern is to auto-label with SAM 3, then fine-tune a smaller model for sub-5ms serving. See the full benchmark comparison.

The flow is three REST calls: request a presigned upload URL and PUT your file, submit a segmentation job with your prompts, then poll the job and download a results archive containing masks and an output manifest. Authentication uses secure API keys with signed requests. Full reference and code examples are in the documentation.