What a run costs (and what the number means)
An estimate of what your provider will bill you, booked per item and per model — visible before you commit, and never a charge from FlowArray.
In this article (5)
FlowArray runs on your own API keys, so every figure it shows is an estimate of what Google, Anthropic, fal.ai or OpenRouter will bill you — not a charge from FlowArray. Your provider invoice remains the source of truth. Within that, the numbers are careful.
Where you see it
- Per item and per run in the run monitor.
- On the confirm button before a regenerate, so you know the price before you press it.
- In the Library, per output.
- Aggregated in Settings → Usage: 7/30/90-day totals, a per-day chart, and breakdowns by model and by person — see usage and model prices.
How each kind is counted
| Kind of step | Counted by |
|---|---|
| Text models | Tokens in and out, at that model's rate |
| Image generation | Per image (about $0.134 for Nano Banana Pro at up to 2K, standard tier; half with Economy) |
| Video (Gemini Omni) | Per second of output — around $0.80 for an eight-second clip at 720p (Veo 3.1 was $3.20) |
| Video (fal.ai — Seedance 2.5 and others) | Per second of output, by resolution — about $9.30 for an eight-second Seedance clip at 1080p, $3.78 at 720p |
| fal.ai cleanup | Per call |
| Voice-over (ElevenLabs) | Per character — about $0.08 per 1,000 |
| Music bed / Sound effect (ElevenLabs) | Per second of audio — about $0.15 / $0.12 a minute |
| Tool calls, GIF, Web-ready video, Extract frames, Edit video, Prompt, Brand guide, Gallery, Send to a URL | $0 — no provider call is made |
Where the prices come from
- OpenRouter-routed models book the actual amount your account was charged for that run, not an estimate.
- Other providers use published list prices, refreshed automatically where the provider offers a feed, and cross-checked against the built-in catalogue so a price change is noticed rather than silently mispriced.
- An admin can override any price in Settings → Usage → Model prices if reality differs; overrides win and take effect within a few minutes.
The honest caveats
- A model that is not in the price list yet shows $0.00. That is a gap in the catalogue, not a free call — the provider still bills you. If you see $0 on a real generation, tell an admin to add a price.
- Some figures are per-call estimates rather than metered (fal.ai background removal is one). Close enough to budget against, not close enough to reconcile a bill with.
- Image pricing assumes the standard tier at up to 1024×1024. Larger sizes, discounted or priority tiers, and responses that mix image and text will drift from the estimate.
- Mock runs book $0 because nothing is called. That is genuinely free — and genuinely not a test of output quality.
Keeping it bounded
Set a daily spend limit on the workspace, and optionally per pool or per key. When the ceiling is reached, new steps stop rather than spending more. This matters most for anything running unattended — a schedule or a trigger spends with nobody watching.