Camera Intelligence — October 2026: Falls & Health vs Everyday Activities

OpenAI gpt-6-luna · TypeSafe AI Jev 1.14.0 · LiquidAI d1-2B (bfloat16) · Qwen3.5-9B (Q4_K_M)

Written by: Agentinosaurus Team
Published: Oct 28, 2026
Reading time: 3 min read

Two floor falls. Three everyday activities. Four AI models, the same frames—and a closer look at when a camera should speak up or stay quiet.

One Event. Five everyday scenes.

A useful camera should notice a fall and stay quiet when someone is simply reading. We asked four models the same question: “A person falls on the floor (health accident).” Each saw the same chronological frames from five staged research clips.

All four detected both floor falls and rejected the three negative cases: 100% agreement and 0% false positives on this small set. Here, “health” means the visible floor-fall Event—not a medical diagnosis. Each image below is one frame from the sequence the models received.

Person falling toward the floor beside a chair — GMDCSA24 / Ekram Alam et al. · Staged research footage

A fall from a chair

The person leaves the chair and falls onto the floor. All four models detected the requested event.

A fall from a chair — results
Model	Answer	Correct	False alert	Response	API cost	RAM¹ GiB	VRAM¹ GiB
OpenAI gpt-6-luna	Match	Yes	—	2.07s	$0.000215	—	—
TypeSafe AI Jev 1.14.0	Match	Yes	—	0.56s	$0.000221	—	—
LiquidAI d1-2B (bfloat16)	Match	Yes	—	2.50s	—	4.25	4.81
Qwen3.5-9B (Q4_K_M)	Match	Yes	—	2.17s	—	6.33	6.65

Person lying across a bed — GMDCSA24 / Ekram Alam et al. · Staged research footage

Landing on a bed

The person lands on a bed, not the floor. All four stayed quiet for our specific Event. This does not establish whether the person needed help.

Landing on a bed — results
Model	Answer	Correct	False alert	Response	API cost	RAM¹ GiB	VRAM¹ GiB
OpenAI gpt-6-luna	No match	Yes	No	1.64s	$0.000215	—	—
TypeSafe AI Jev 1.14.0	No match	Yes	No	0.27s	$0.000221	—	—
LiquidAI d1-2B (bfloat16)	No match	Yes	No	0.85s	—	4.25	4.81
Qwen3.5-9B (Q4_K_M)	No match	Yes	No	2.13s	—	6.33	6.65

Person seated on a bed with material in their lap — GMDCSA24 / Ekram Alam et al. · Staged research footage

Sitting on a bed

The person remains seated on the bed. All four models correctly returned no match.

Sitting on a bed — results
Model	Answer	Correct	False alert	Response	API cost	RAM¹ GiB	VRAM¹ GiB
OpenAI gpt-6-luna	No match	Yes	No	1.28s	$0.000267	—	—
TypeSafe AI Jev 1.14.0	No match	Yes	No	0.67s	$0.000277	—	—
LiquidAI d1-2B (bfloat16)	No match	Yes	No	0.79s	—	4.25	4.81
Qwen3.5-9B (Q4_K_M)	No match	Yes	No	2.22s	—	6.33	6.65

Person reading while seated on a bed — GMDCSA24 / Ekram Alam et al. · Staged research footage

Reading quietly

A person reads while sitting on a bed. All four models correctly stayed quiet.

Reading quietly — results
Model	Answer	Correct	False alert	Response	API cost	RAM¹ GiB	VRAM¹ GiB
OpenAI gpt-6-luna	No match	Yes	No	0.28s	$0.000267	—	—
TypeSafe AI Jev 1.14.0	No match	Yes	No	0.39s	$0.000277	—	—
LiquidAI d1-2B (bfloat16)	No match	Yes	No	2.60s	—	4.25	4.81
Qwen3.5-9B (Q4_K_M)	No match	Yes	No	2.34s	—	6.33	6.65

Person on the floor next to a chair after a fall — GMDCSA24 / Ekram Alam et al. · Staged research footage

A second fall onto the floor

Another seated fall reaches the floor. All four detected it. This clip shares a subject and room with the first fall.

A second fall onto the floor — results
Model	Answer	Correct	False alert	Response	API cost	RAM¹ GiB	VRAM¹ GiB
OpenAI gpt-6-luna	Match	Yes	—	0.40s	$0.000190	—	—
TypeSafe AI Jev 1.14.0	Match	Yes	—	0.58s	$0.000193	—	—
LiquidAI d1-2B (bfloat16)	Match	Yes	—	0.62s	—	4.25	4.81
Qwen3.5-9B (Q4_K_M)	Match	Yes	—	1.82s	—	6.33	6.65

Same answers. Different response times.

TypeSafe was fastest in this run: 0.56s median, followed by Liquid at 0.85s, OpenAI at 1.28s and Qwen at 2.17s. Accuracy was tied on these cases. Estimated total cloud usage was $0.00115 for OpenAI and $0.00119 for TypeSafe. Local inference has no provider fee; hardware and electricity still cost money.

¹ Memory figures show peak usage on an RTX 4080 Laptop GPU, including model loading. Liquid and Qwen use different GPU memory measurements, so these figures are approximate comparisons, not minimum hardware requirements.

Reading the results

All models saw the same images. These five staged scenes are a small comparison, not an estimate of performance in every home or workplace.

A match means the requested fall was detected. A false alert means an alert on a scene without that Event.

Response times measure the model response, not the time from an incident to an alert. Cloud costs are estimates; local hardware and electricity are excluded. “—” means unavailable or not applicable.

Written by

Agentinosaurus TeamThe people building AgentinosaurusView profile

We build Agentinosaurus, an independent product that gives the cameras you already own an Agent that pays attention. You describe what matters in plain words, and it tells you the moment it happens.