Meta Launches SAM 3.1 API: Detection, Segmentation and Tracking in One Call
On September 19, Meta launched Segment Anything Model 3.1 (SAM 3.1) on the Meta Model API. With a single noun-phrase prompt, developers can locate targets in images and videos; one API call returns detection results, pixel-level segmentation masks, and identity-consistent video tracking. The API is positioned as lightweight, fast, and low-cost, and has already sparked community discussion of automation use cases.
Confirmed
- SAM 3.1 is officially available on the Meta Model API as an official implementation with inference optimizations tailored to the model architecture (relayed by @armandruiz).
- Core capability: prompt with a noun phrase to detect, segment, and track objects in images and videos; a single API call returns detection results, pixel-level masks, and identity-consistent video tracking (consistently relayed across multiple posts).
- Pricing: relayed by @Vjeux as $2.5 per 1,000 images.
Why it matters
- SAM 3.1 merges detection, segmentation, and video tracking into a single call, cutting the cost of multi-stage vision pipelines for speed- and cost-sensitive applications.
- @threepointone noted that its faster, more accurate, cheaper object segmentation opens up composable possibilities, such as pairing it with browser automation to build fully autonomous Computer Use agents.
2026-09-19 ~ 2026-09-19 · 5 related posts
Primary sources
- Meta ships SAM 3.1 on Model API: noun-phrase prompts for detection, segmentation and tracking at $2.50/1k images — armand_ruiz ·
- Meta ships SAM 3.1 API: detect, segment and track with a noun phrase, $2.50 per 1,000 images — Vjeux ·
- SAM 3.1 Arrives With Faster, Cheaper, More Accurate Object Segmentation — threepointone ·
- Meta ships SAM 3.1 on Model API: detection, segmentation, tracking in one call — AIatMeta · 2026-09-19
- [source] Meta ships SAM 3.1 on Model API: noun-phrase prompts for detection, segmentation and tracking at $2.50/1k images — armand_ruiz · 2026-09-19
- [source] SAM 3.1 Arrives With Faster, Cheaper, More Accurate Object Segmentation — threepointone · 2026-09-19