Mercury caught 93.5% of fall images, compared with just 70.2% for Claude.
Loading the verified results from the completed benchmark.
Mercury devices only. These are image-level classification results, not independent fall-event detection or alert performance.
Real-time person activity and emergency monitoring with a Ring camera, powered by Mercury Alert AI’s custom model.
See Mercury Instant’s activity-monitoring interface in action. This product recording is separate from the benchmark evaluation below.
View on the Ring AppstoreFive states. The same images for both models.
K-means clusters. Nearest-neighbor representatives.
The reference evaluation contains 15,000 images selected from ten camera views. Its public release sample contains 1,151 images, with up to 25 per camera and reference state.
Within each camera and state, K-means uses k equal to the allocated image count. One actual image nearest each cluster center is selected. The release sample is selected directly within the frozen reference set.
Neither model’s predictions nor correctness influenced image selection. These are deliberately selected camera views, not a random sample of homes.
Counts describe labeled images, not independent incidents. Room numbers in the explorer identify individual camera views.
More than one number.
Performance by position
Same images. Same reference labels.
Where the labels went
Rows: reference label · columns: prediction
Cell values are image counts. Shading shows the share of each reference-labeled class.
Every agreement. Every disagreement.
Select a group to inspect its images.
Follow the similarities.
Drag or scroll through one continuous globe containing the whole shared sample. Images grow as they come toward you. Select one to inspect both models’ predictions.
Choose a room
Export these rows ↓Room numbers identify camera views, not verified unique rooms. Select one to explore images from that view; position and result filters still apply.
— matching images · Explorer filters do not change the headline benchmark metrics.
Even spacing · Turn on to show approximate similarity distances on the globe.
Drag or scroll to explore the matching images; thumbnails load as they come into view. Filtering regroups the selected images into a compact globe. Reset filters to restore the full collection. Images grow as they approach the middle. Cosine-similarity spacing is an approximate browsing layout.
Every Mercury camera, measured.
Inspect where performance changes, with the denominators visible.
| Room | Images | Fall images | Mercury Instant | Claude | Difference |
|---|
Camera sources are not necessarily different rooms. Images from the same recording may be strongly related. Unique rooms, fall incidents, and posture scenarios have not been independently annotated or counted in this run.
Built for continuous monitoring.
Inference times and estimated serving costs for Mercury Instant and Claude Opus 5.