Skip to content

Pricing

Agnify is pay-as-you-go. You’re billed only for the inference that actually runs — no seats, no subscriptions, no per-minute video tax. Every new account starts with $5 of free credit, and you top up from there whenever you want.

Plug in a video length, how many frames per second you want analyzed, and which nodes your pipeline uses. The estimate updates as you type.

Nodes in your pipeline
$0.000001/frame
$0.000028/frame
$0.000028/frame

Pose already includes detection — no need to also tick D-FINE for the same objects. Frame description prices by the model you pick — a self-hosted Qwen model is far cheaper than the Gemini API. Attribute extraction bills per finished track, not per frame, so it isn't estimated here.

0 frames analyzed per node
    Estimated cost $0.00

    Pricing is per successful inference call. One “call” is one priced node processing one analyzed frame. So the cost of a run is:

    cost = Σ (rate for each node) × (frames analyzed by that node)

    A few consequences worth knowing:

    • Only frames you analyze count. A 30 fps video analyzed at 1 fps is metered as 1 frame per second, not 30. You set the sampling rate on the FrameExtractor node.
    • Only successful calls are billed. Failed, mocked, or empty-input calls are never metered. A run that crashes on its first wave costs almost nothing.
    • Unpriced nodes are free. Frame extraction, tracking, counting, redaction, and output nodes don’t call inference, so they add nothing to your bill.
    • Detections returning zero objects still count. A successful inference that finds nothing is still a call that ran on a GPU.
    • Pose estimation includes detection. Pose runs as a single fused detect-and-pose call, so a pose pipeline is billed at the pose rate alone — you aren’t charged separately for object detection on the same frames.
    • Attribute extraction is billed per finished track, not per frame. It makes one VLM call for each tracked object as it leaves the scene — so its cost depends on how many distinct objects you see, not on frame count. That’s why it isn’t in the frame-based calculator above.
    • The VLM nodes price by the model you choose. Frame description and attribute extraction can run on a small self-hosted model or the frontier Gemini API, and you’re charged the real cost of whichever you pick — a Gemini call costs roughly 30× a self-hosted 8B call.

    Rates are zero-margin — self-hosted rates reflect measured GPU cost, and the Gemini rates pass through Google’s API price, with no markup.

    NodeWhat it doesBilledRate
    Object detection (D-FINE)Bounding boxes for objectsper analyzed frame$0.000001
    Segmentation (SAM3)Pixel masks for objectsper analyzed frame$0.000028
    Pose estimationBody keypoints (detection included)per analyzed frame$0.000028

    The VLM nodes — Frame description (per analyzed frame) and Attribute extraction (per finished track) — price by the selected model:

    ModelHostingRate per VLM call
    Qwen3.6 35Bself-hosted$0.000056
    Qwen3.5 9Bself-hosted$0.000028
    Gemini 3 FlashGoogle API$0.000380
    Gemini 3 ProGoogle API$0.001550

    For example, a 5-minute video analyzed at 1 fps (300 frames) with object detection plus frame description on the Qwen3.5 9B model costs 300 × ($0.000001 + $0.000028) = $0.0087; switch frame description to Gemini 3 Pro and it’s 300 × ($0.000001 + $0.001550) = $0.4653. The same clip with pose estimation (a fused detect-and-pose call) is 300 × $0.000028 = $0.0084.

    Your balance is a prepaid credit ledger, denominated in US dollars.

    • Free signup credit: every new account gets $5 to start, automatically.
    • Top up any time from the Top up menu in the app header or from Billing in organization settings. Preset amounts are $10, $25, $50, and $100, paid by card through Stripe’s hosted checkout.
    • Runs start while you have a positive balance. When your balance hits zero, new runs are blocked until you top up — but an in-flight run is never killed mid-way, so your balance can dip slightly negative as the last charges settle.