Pricing and Limits
Understand Cloud Streaming credit usage, GPU sizing, price estimates, and processing limits before starting a run.
Cloud Streaming bills for the hours a Workflow runs on a GPU, at a rate set by the GPU size. Previews, tests, persistent deployments, and uploaded-video processing all use credits. The credits page lists the rate for each GPU size. Your Workspace's rates can differ (ex: on an Enterprise contract), and the app shows your rate before you start.
GPU Sizes
Small
gpu-s-v1
Object detection (ex: RF-DETR, YOLO26)
Medium
gpu-m-v1
Small vision language models (ex: Qwen3-VL 2B)
Large
gpu-l-v1
SAM 3
A larger GPU costs more per hour. The app and the Roboflow Agent pick the size from the Workflow's models, so you don't choose it. The HTTP API can request a size.
Web App
Review the estimate
The standalone source page shows an estimated hourly rate before you start. "Use Workflow", then "Cloud Streaming", can also show the saved Workflow's estimate. An estimate does not reserve capacity or lock a price or Workflow revision.
"Start Processing" uses the saved Workflow resolved at start time. If you edit it after viewing the estimate, the starting allocation and rate can differ. A running job keeps its own Workflow snapshot; later edits do not update it automatically.
The standalone source page shows the selected run's current allocation rate, including runs started by the Agent. This is not accumulated spend. Even on a finished job, a returned processing rate reflects current Workspace pricing, not the historical bill for that run.
Normal Agent conversations, editor previews, stream lists, and shared output views do not show the rate. Promotion confirmation is an exception: it shows the running preview's current rate before you make it continuous. These display differences do not change billing.
Automatic sizing and price changes
App and Agent starts use automatic GPU sizing. The initial size is based on the Workflow; it is not a frame-rate guarantee. When an eligible run has a confirmed GPU-memory failure, the platform can stop that attempt and retry the same Workflow snapshot on a larger GPU, subject to capacity and credit checks.
While resources are being adjusted, wait for the replacement run and its updated rate. Running jobs are not automatically downsized. If the Workflow cannot fit on the largest supported allocation, reduce its resource needs, such as by using a smaller model.
Promotion keeps the running preview's allocation and rate. It changes its lifetime, not its resource size. See Run a Workflow for preview, deployment, restart, and promotion behavior.
Credit usage
Processing usage is metered by reported processing time and GPU allocation, not by how many viewers are watching. Workflow model and block records do not create a second Cloud Streaming processing charge. Other billable services used by the Workflow or Agent retain their own billing rules, listed on the credits page.
New starts, restarts, and automatic size increases check whether the Workspace can spend credits. Running jobs also check credit standing and stop at the credit limit. A credit-limit stop stays stopped rather than automatically recovering after a refill; resolve billing and start or restart explicitly.
Open Workspace Settings, then "Credit Usage", to review actual consumption, date ranges, usage types, and API-key attribution where available. See View Your Credit Usage. The source page's hourly rate and Activity panel are not billing ledgers.
Availability and access
Cloud Streaming is publicly available in all environments. You need access to the source and Workflow and sufficient credits for processing.
"View Stream Processing" grants read access. Starting, stopping, watching, and managing sources need "Manage Source Video" or the corresponding API authority. Saved captures also require upload permission. See permissions.
Processing limits
A persistent deployment has no fixed product duration cap. Keep it running only as long as needed, then stop it explicitly.
A draft preview is bounded and is not an unattended deployment. Browser webcam and phone sessions cannot be promoted to persistent processing. See preview lifetime.
Concurrent placement depends on available capacity, source placement, and Workspace policy. There is no guaranteed start time or frame rate.
Workflows must fit the supported GPU resources and use blocks compatible with their cloud runtime. Verify your model and block combination with a controlled preview before deployment.
Source capture FPS, processed FPS, and browser playback FPS measure different stages. A requested FPS cap is a limit, not a promise that every camera frame is processed or displayed.
Video playback depends on the camera codec, browser, and network. JSON-only Workflows do not require an annotated image output. Browser MP4 capture has its own support and lifetime requirements in Consume Results.
Live results are bounded and temporary. For durable selected results, use a sink; for finite uploaded-video artifacts, see Run a Model on a Video.
HTTP API
A job you start with the API uses the GPU size you send, or Small (gpu-s-v1) if you send none. Unlike app starts, it is not sized from the Workflow and does not move to a larger GPU after a GPU memory error. Do not send app-only sizing or paid-validation fields.
Request field limits, authentication, and response shapes are defined in the job operation. A request limit such as maxFps is not a throughput guarantee. See Troubleshooting for credit, capacity, and resource failures.
Last updated
Was this helpful?