Scaling AI Video Generation: Batching, Webhooks and Cost Control

The Nureta Developer API prices video linearly, so cost is seconds × rate. seahorse-480p: $0.1888 per second, $0.944 per 5 seconds. seahorse-720p: $0.2773 per second, $1.3864 per 5 seconds. seahorse-1080p: $0.5428 per second, $2.7139 per 5 seconds. Single clips run 5, 8, 10, 12 or 15 seconds; stitched multi-shot videos run any whole number of seconds from 16 to 180 on the same curve. Scripts cost a flat $0.01 and images a flat $0.05. The public endpoint GET /api/v3/models returns current prices, so read rates from it instead of hard-coding them.

These are API prices, not app prices. The consumer app at nureta.ai charges $1.29 per 5-second video, or 2 credits per second, which works out to about $0.91 per 5 seconds on the $39.99 bundle. The app adds a free first video and a finished interface at 480p; the API is a metered developer product with 720p and 1080p, billed from a prepaid balance. Treat them as different products at different prices.

Worked examples at list prices. One 60-second film: $11.33 at 480p, $16.64 at 720p, $32.57 at 1080p. One 180-second film at 480p: $33.98. A batch of 100 clips of 15 seconds at 480p: 1,500 seconds × $0.1888 = $283.20; the same batch costs $415.95 at 720p and $814.20 at 1080p. 100 storyboard scripts: $1.00. 100 cover images: $5.00.

The lesson is that resolution, not script or image volume, is the main cost lever. Moving from 480p to 1080p nearly triples the price of every second, while scripts and images are close to free next to video. Budget a run as total seconds per resolution, then add a small allowance for re-rendering shots you reject.

Throughput is set by a per-key rate limit. New keys default to 10 requests per second, adjustable from 1 to 100 per key. Every call counts, not just task creation: status polls, storyboard edits and upload requests draw from the same budget. Upload presigning also has a fixed cap of 30 requests per minute per key, so upload reference images before a large batch starts, not inside the hot loop.

Over the limit, the API returns HTTP 429 with the code RateLimitExceeded and standard RateLimit headers. Read the headers, back off briefly and retry. A queue that caps your own send rate slightly under the key limit avoids most 429s, and separate keys for separate workloads, such as bulk renders versus a live product, stop a batch from starving interactive users.

Polling a thousand running tasks every few seconds spends your rate limit on answers that mostly say nothing changed. Instead, set a callback_url on each task and the API POSTs the final task object to you when it succeeds or fails, the same object a GET returns. Add a callback_secret of 8 to 256 characters and every delivery is signed with HMAC-SHA256, which you verify with a constant-time comparison.

Build the receiver for the delivery rules. Delivery is at-least-once, so deduplicate by task id. A delivery succeeds on any 2xx response within 5 seconds, redirects are not followed, and failures are retried for up to 5 attempts, so acknowledge fast and download the video in a background job. The URL must be HTTPS and cannot target private or localhost addresses. Keep a slow fallback poll for tasks that never report back.

The prepaid balance is a hard cap. Creating a task reserves its price up front, and if the balance cannot cover it the API returns AccountOverdueError instead of running the job, so a runaway loop cannot spend money you have not loaded. Failed generations are refunded automatically, so you pay only for delivered videos. Load a balance per campaign and the batch cannot overshoot it.

Three habits cut waste further. Test ideas with a $0.01 script pass, which returns the storyboard without rendering. Use edit mode as an approval gate for long videos: the draft parks with preview images, and cancelling before render refunds the full charge immediately. And prototype at 480p: resolution is fixed by the model chosen when a task is created, so once a storyboard is approved, submit it as your own segments in a new task on seahorse-720p or seahorse-1080p. The storyboard controls themselves are covered in our guide to storyboard-level control.

Every video task response carries a step object with a stable machine code such as queued, generating_script, generating_visuals, pending_rendering, rendering, stitching, done, failed or cancelled. While rendering it also counts finished clips, for example {"code": "rendering", "done": 3, "total": 5}. Map the codes to your own dashboard labels, and expect the step to lag the true phase by up to about 30 seconds.

Useful batch metrics come straight from this data: tasks per status, seconds rendered per resolution, spend reserved versus refunded, and time from creation to success. Watch the refund rate and the share of drafts you cancel; both show whether prompts need work before the next run. Queues, storage and pipeline design are covered in our guide to automating adult video production with an API.