How per-second pricing works for realtime video, with worked examples
Understand billable-second video API pricing, compare model rates and estimate session costs. Queueing and startup are free; release sessions to stop billing.
Windflow bills realtime video by the billable session-second: one metered second after generation starts, excluding documented reconnect grace. Rates run from $0.0017 to $0.0069 per second depending on the model, which is $6.12 to $25.00 per hour of billable time. There is no subscription, no seat fee and no minimum.
This guide shows what is and is not metered, what a session costs on each model, and how to estimate a product's bill before you build it.
The price table
Current public rates, in US dollars. They are indicative private-beta rates and are also exposed at GET /v1/pricing.
| Model | Per billable second | Per minute | Per hour | Ten minutes of billable usage |
|---|---|---|---|---|
| Helios | $0.0017 | $0.102 | $6.12 | $1.02 |
| LongLive-2.0 | $0.0021 | $0.126 | $7.56 | $1.26 |
| SANA-Streaming | $0.0019 | $0.114 | $6.84 | $1.14 |
| EdenLive | $0.0021 | $0.126 | $7.56 | $1.26 |
| JoyAI-Video-Edit | $0.0021 | $0.126 | $7.56 | $1.26 |
| LingBot-World 2.0 | $0.0024 | $0.144 | $8.64 | $1.44 |
| Matrix-Game 4.0 | $0.0024 | $0.144 | $8.64 | $1.44 |
| DreamX-World-5B | $0.0023 | $0.138 | $8.28 | $1.38 |
| ABot-World-0-5B | $0.0023 | $0.138 | $8.28 | $1.38 |
| MiniMax FastH3 | $0.0069 | $0.417 | $25.00 | $4.17 |
| LTX-2.5 | $0.0040 | $0.240 | $14.40 | $2.40 |
| Hy-Nomas | $0.0047 | $0.283 | $17.00 | $2.83 |
These are usage budgets, not a promise of one uninterrupted ten-minute session. Model-specific session ceilings still apply; the session response states session_max_seconds.
What counts as billable time
Billing starts at the session's first generation_started event and stops when the session is released or ends. Queueing, worker startup, time before generation starts and the documented reconnect grace are not billed. After generation starts, paused or unwatched time remains billable; release the owner session to stop the meter.
- Not billed: time in the queue, worker startup, time before generation starts, and the documented reconnect grace.
- Billed: every billable second, including seconds when nobody is watching the stream.
- Pausing the video does not stop the meter. Pausing a track only stops delivery to that connection, and a paused generation can still hold the GPU. Release the session when the experience is finished or idle.
- Extra viewers are free. One session is metered once, however many authorised guests receive it. Separate concurrent sessions each carry their own meter.
Why the rates differ
Rates follow the GPU and runtime profile a model needs to stay interactive, not an abstract unit such as pixels or output seconds. A 5-billion-parameter world model and a 35-billion-parameter audio-video model do not occupy the same hardware, so they do not cost the same per second.
Estimating a product's cost
Cost is sessions times seconds times rate. Two worked examples:
- A demo on a landing page. 200 visitors each run a 3-minute Helios session: 200 × 180 s × $0.0017 = $61.20.
- A creative tool. 20 creators each steer LingBot-World 2.0 for 2 hours a day: 20 × 7,200 s × $0.0024 = $345.60 per day.
Two habits keep the number honest. Count seconds from the first generation_started event until release or session end, excluding documented reconnect grace. Then add the billable time of each independent owner session. Peak concurrency determines the simultaneous capacity you need; it is not an extra multiplier if your total session count already includes those users.
A session timeline to include in your budget
| Stage | Metered? | What your application should do |
|---|---|---|
| Queue and worker startup | No | Show the actual waiting state; do not promise immediate capacity |
| Connected and ready, before first generation | No | Stage the model's required inputs and start deliberately |
| Generation has started | Yes | Count elapsed billable time at that model's rate |
| Playback is paused or nobody is watching | Yes, after generation starts | Release an owner session that no longer needs to retain capacity |
| Documented reconnect grace | No | Reconnect within policy, or end the experience clearly |
| Session released or ended | No further session time | Confirm cleanup; do not rely only on hiding the player |
A hardware benchmark cannot replace this timeline. For example, the EdenLive compute measurements count model-loop throughput rather than hosted-session billable time. Compare total billable usage and your required concurrency against the actual cost of dedicated capacity, including idle time and the delivery system you would operate.
What a $5 grant buys
Approved accounts start with $5 and four concurrent sessions. At the published rates that is about 49 minutes of Helios, 21 minutes of LTX-2.5 or 12 minutes of MiniMax FastH3. It is enough to evaluate several models for real, not just read about them.
Why you may see a 402 with a positive balance
Creating a session reserves the model's rate multiplied by its maximum session length. Concurrent reservations reduce the balance you can spend. When a session ends, the actual billable seconds settle and the unused part of the reservation returns to your balance.
How to keep the bill down
- Release sessions the moment a user leaves. After generation starts, an idle session still bills until it hits the model's session ceiling.
- Prototype on the cheaper models. Helios, SANA-Streaming and the world models cost under $0.0025 per second; move to LTX-2.5 or FastH3 when you need generated audio.
- Share one session. A broadcast to many viewers is one meter, not many.
- Cap session length per user. Short-lived, scoped tokens can carry a session budget, so the worst case a user can spend is known before the token leaves your server.
Renting a GPU versus paying per session
Renting a GPU bills the whole hour whether or not anyone is using it, and leaves you to build the streaming layer, the session state and the scheduling. Per-session billing charges only billable seconds and includes those parts. For steady, round-the-clock load on one model, dedicated capacity can be the better fit; Windflow scopes that separately as reserved scale. For spiky, interactive or early-stage traffic, per-second billing keeps cost proportional to use.
FAQ
What is the cheapest way to run a realtime video model without managing GPUs?
Pay per billable session-second on a hosted platform. On Windflow the lowest rate is $0.0017 per second (Helios), which is $6.12 for an hour of billable time, with no GPU to provision and no charge while a session is queued or starting.
Is there a free trial?
Approved accounts start with $5 of usable balance across the public models.
Are recordings and uploads billed separately?
During private beta the published billable-session rate is the meter. Temporary recordings expire after 24 hours; high-volume exports, dedicated storage and private delivery are scoped separately.
Where can I see current rates?
On the pricing page, in each model's page, and at GET /v1/pricing.