September 25th, 2026 Model
Introducing Loop 1.5
A New Layer of Intelligence for Embodied Agents
Today we're releasing Loop 1.5: a model built to control embodied agents. It is our most performant public model, adding native audio understanding, video object tracking, web search, sub-agent calls, and stronger visual reasoning to the Loop family.
Equally important, Loop 1.5 powers agentic applications across many physical embodiments, from drones and quadrupeds to smart glasses and phones, all without platform-specific retraining.
Tracking an object through a video
To complete real-world tasks such as inventory tracking or sports analytics, an agent must follow unique entities across time. Loop 1.5 emits object tracks as timestamped geometries instead of one-off per-frame detections, which removes entire re-identification pipelines downstream.
The result is state-of-the-art tracking accuracy on the video object segmentation benchmarks we measure, whether you are counting stock on a retail floor or following players on a broadcast feed.
Understanding egocentric video
First-person video unlocks everything from annotating robotic training data to helping smart glasses interpret what a wearer sees. Trained specifically on egocentric footage, Loop 1.5 localizes hands and segments subtasks far better than the previous generation.
Adding audio as a new modality
Loop 1.5 natively understands audio: time-aligned transcription, sound-based event detection, and audio-visual question answering work out of the box, no separate pipeline required.
Learning to use tools
Loop 1.5 calls arbitrary tools declared in OpenAI function format. It decides what to call and when, including web search, page reads, zoom, and reverse image search — knowledge it never had to memorize in training becomes available at runtime.
Faster inference while pushing the frontier
Loop 1.5 completes end-to-end requests up to 4x faster than Loop 1.0 while moving accuracy forward on video understanding and embodied reasoning benchmarks.